Survey of Wireless Network-on-Chip Systems

by

Xi Li

A report submitted to the Graduate Faculty ofAuburn University

in partial fulfillment of therequirements for the Degree ofMaster of Electrical Engineering

Auburn, AlabamaMay 10, 2012

Keywords: wireless, NoC

Copyright 2012 by Xi Li

Approved by

Vishwani Agrawal, Chair, James J. Danaher Professor of Electrical and ComputerEngineering

Shiwen Mao, Associate Professor of Electrical and Computer EngineeringJitendra Tugnait, James B. Davis and Alumni Professor of Electrical and Computer

EngineeringMark Nelms, Professor and Chair of Electrical and Computer Engineering

Abstract

Nowadays, network-on-chip (NoC) systems are becoming more popular due to their big

advantages when compare with systems-on-chip (SoC). Therefore, an increasing number of

researchers and organizations now focus on the study and development of NoC techniques.

As a result, so far many achievements have been gained. Furthermore, considering the

dominant position of wireless and the weakness of wired communication, people also turn to

try to insert wireless links in NoC systems in order to solve the multi-hop problem.

This report gives a brief description of some outstanding developments of NoC and

WNoC (Wireless NoC) systems, including some important technique and the achieved re-

sults, mainly related to the required hardware and communication protocols. In addition,

the report also contains my experiments on NoC and WNoC systems. I use the Booksim

simulator to measure their performances and make some comparisons, and then give some

analysis and conclusion of those results. At the end the report summarizes the finished work

and gives some more developing directions of NoC and WNoC systems.

ii

Acknowledgments

I would like to express my gratitude to people who helped me in this project.

The first person I must thanks is my advisor Dr. Vishwani D. Agrawal, the James J.

Danaher Professor at Electrical Engineering Department. During the process of the project,

he lent me so much help. When I met problems, he would guide me to solve those problems,

either through emails or in face to face meetings, he always helped me with lot of patience.

Dr. Agrawal’s kindness and patience made me become interested in this project and provided

me much energy to do my work and also gave me a hope when I felt frustrated.

I am also deeply grateful to my committee members, Dr. Shiwen Mao and Dr. Jitendra

Tugnait for their great teaching and patience during my study.

Another two people I must thanks to are Dr. Alireza Babaei at Auburn and Dr. Partha

Pratim Pande from Washington State University. Dr. Babaei helped throughout the project.

Dr. Pande offered me much useful references. From his help and references, I gained a

significant amount information that I needed.

I greatly appreciate my colleague and friend, Suraj Sindia, for his help about how to

use latex easily.

Last but not least, I must thank my parents. Though they are no experts in this area,

they encourage me so mush that when I meet difficulties, no matter how big they are, I never

lose hope and at last I can overcome them.

Many thanks to my dearest friends in Auburn. Without them, my life in Auburn could

not have been so wonderful.

iii

Table of Contents

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii

Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Problems in Traditional NoC . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Some Solutions and Motivation for Studying Wireless NoC . . . . . . . . . . . . 6

3.1 Ultra Wide Band (UWB) Based WNoC System . . . . . . . . . . . . . . . . 7

3.2 A mm-Wave WNoC System . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3.3 CNT Based WNoC System . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3.4 Other Needed Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

4 Technique Used in a WNoC System . . . . . . . . . . . . . . . . . . . . . . . . . 12

4.1 Structure and Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4.1.1 Structure of a Pure Wireless NoC System . . . . . . . . . . . . . . . 12

4.1.2 Structure of Hybrid Wireless NoC System . . . . . . . . . . . . . . . 13

4.2 Wireless Link Insertion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

4.3 Routing and Communication Protocol . . . . . . . . . . . . . . . . . . . . . 14

4.3.1 Protocols in Pure Wireless NoC System . . . . . . . . . . . . . . . . . 15

4.3.2 Protocols in a Hybrid Wireless NoC System . . . . . . . . . . . . . . 17

5 A Simulator for NoC System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

6 Factors and Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

7 Experimental Processes, Results and Analysis . . . . . . . . . . . . . . . . . . . 25

7.1 Experimental Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

iv

7.2 Experimental Result and Analysis . . . . . . . . . . . . . . . . . . . . . . . . 28

7.2.1 Latency vs. Injection Rate . . . . . . . . . . . . . . . . . . . . . . . . 28

7.2.2 Latency vs. Virtual Channels . . . . . . . . . . . . . . . . . . . . . . 36

7.2.3 An Analysis of WNoC System . . . . . . . . . . . . . . . . . . . . . . 38

8 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

8.1 Work Completed in This Report . . . . . . . . . . . . . . . . . . . . . . . . . 40

8.1.1 Device-Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

8.1.2 System-Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

8.1.3 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

8.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

v

List of Figures

7.1 Example configuration file for simulating a mesh NoC system. . . . . . . . . . . 26

7.2 Simulator output from running the examples/mesh88 lat configuration file. . . . 27

7.3 A 4× 4 mesh topology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

7.4 Latency vs. injection rate in mesh topology. . . . . . . . . . . . . . . . . . . . . 29

7.5 A 4× 4 mesh topology with concentration 4. . . . . . . . . . . . . . . . . . . . . 30

7.6 Average hops in 8× 8 mesh. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

7.7 Average hops in 8× 8 mesh with concentration 4. . . . . . . . . . . . . . . . . . 31

7.8 Latency vs. injection rate in cmesh topology. . . . . . . . . . . . . . . . . . . . 32

7.9 Comparison between regular mesh and concentrated mesh. . . . . . . . . . . . . 33

7.10 (a) Flattened butterfly topology consisting of 64 nodes; (b) corresponding router

layout. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

7.11 Latency vs. injection rate in flattened butterfly topology consisting of (a) 64

nodes and (b) 256 nodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

7.12 Simulated latency vs. number of virtual channels for (a) regular mesh and (b)

flattened butterfly topologies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

7.13 Comparison between regular mesh and flattened butterfly with (a) 4 virtual chan-

nels, (b) 8 virtual channels and (c) 16 virtual channels. . . . . . . . . . . . . . . 39

vi

List of Tables

3.1 WNoC system parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

vii

Chapter 1

Background

Many publications [4], [28], [55] state that the emergence of SoC (System-on-Chip)

platforms consisting of a larger number of embedded processors is a outstanding solution

for some industrial tasks. As described in [23], [76], an SoC is a type of micro system that

integrates many components like processor cores, DSP cores, memories (or storage of control

interface that may be out of the chip) and many other hardware cores, which otherwise

perform specialized tasks on separate single dies. However, according to [23], [65], [76], there

are two limitations of SoC: first of all, the communication among the Intellectual Property

(IP) blocks impedes the development at the system level, aggravating the complexity issue.

The global wire delays are the most important factors that typically do not scale with

technology scaling even after we insert repeaters [29]. Another problem with SoC is that, SoC

always integrates several different hardware cores for different applications on the same chip,

but those applications always have diverse requirements such as the different communication

standards and specific design constraints. Therefore, there may be difficulties when designing

a common SoC for many applications.

In order to solve the problems stated above, NoC (Network-on-chip) is a good paradigm

[3], [4], [23], [64], [76]. NoC is an integrated network that uses routers to allow the commu-

nication among those blocks. It makes use of networking theory and methods for on-chip

communication so that the blocks can exchange information on a chip just like what the

terminals do in the actual world. The so called blocks here will always refer to processors

and caches and, for simplify, a block will be renamed as node in the remaining report.

In a system, the distribution of nodes complies with certain specific topologies. Some

popular topologies have been studied, such as SPIN (Scalable, Programmable, Integrated

1

Network), CLICH (Chip-Level Integration of Communicating Heterogeneous Elements), reg-

ular mesh, Torus, Fold torus, Octagon and BFT (Butterfly Fat Tree) [26], [65]. Each node

communicates through routers and since there must be a numbers of routers integrated on

a centimeter-size chip, the size of router should be small enough and the structure should

be as simple as possible. In general, a router in NoC system consists of 1) buffers to store

data temporarily, 2) arbiters to decide the sequence of data transmission and 3) switches to

transfer data in the right direction. So far, the wormhole router is the most popular router

used in NoC systems. The advantage of this type of router is that data can be transmitted

more fluently. Corresponding to the wormhole router, the transmitted data unit in NoC

is usually called “flit”. A flit contains dozens of bits. Each flit has a header, which con-

tains the address of the destination, and it is followed by a string of data bits. When being

transmitted, the flit runs like a stream through the routers.

A link in NoC is point to point. The common communication between two nodes

is generally based on packet-switching, although there exist other NoC proposals utilizing

circuit-switching techniques. As each channel that connects nodes is duplex and shared by

only two nodes, the NoC improves the performance over the traditional system, which uses

shared-bus to transmit data.

In a NoC system, time latency, power consumption and throughput are the main pa-

rameters used to evaluate its performance. Latency is defined as the time (in clock cycle)

that elapses between the occurrence of a message header injection into the network at the

source node and the reception of the tail of message at the destination node [63]. Therefore,

for a given message, the latency L is given by [65]:

L = sender overhead+ transport latency + receiver overhead (1.1)

The transport latency is tiny so that we can ignore it. Send overhead and receiver

overhead mainly depend on the message waiting time in a router, which is in proportion

to the number of nodes on a single chip times the injection rate. For example, if a node

2

requires to transmit data to another node, the data must pass through all routers along the

path between those two nodes. Although a path can be readily determined by any kinds of

routing protocol, the transmitted data must encounter every router on the path. Assuming

a standard mesh network, if node A, positioned is at (1, 1), intends to transmit data to node

B at position (N,M), where (N > 2, M > 2), with some high probability there can be other

transmitting paths that cut across the long path between nodes A and B. That means the

routers in the path from node A to node B will be used by other transmissions at the same

time. Since a router can process only one transmission at a time, when several transmissions

come to the same router, only one of them can be processed immediately while the others

wait until the router is free. In such cases, a high latency in the NoC system is encountered.

Power consumption is another parameter that affects the performance of an NoC system.

Since there are many node-to-node hops in a path and each hop consumes some energy, the

multi-hop must cost more total energy. For each hop, router is the main source of the power

consumption. When a flit passes by a router, it will be stored in an input buffer to wait

for being processed. If there are more than one flits stored in different input-buffers, an

arbiter should decode the destination information of each flit and decide which flit to send

first according to the routing protocol being used. Then, the chosen flit is transmitted from

input-buffer to output-buffer through a crossbars switch. If there is only one flit arrives at

the output of the router, it can be sent directly. But if there are several flits they have to

be stored in the output buffer temporally until their turn. The power dissipated in a NoC

can be calculated as follows [44]:

P =∑

Pr buf + Pr arbiter + Pr crossbar + Pr link (1.2)

Pr buf is the average power consumption in buffers including both dynamic and static

power; Pr arbiter is the average power consumption in routing computation; Pr crossbar is

the average power consumption in switches used in data transmission from input to output

3

through a router; and Pr link is the average link power consumption between two neighbor-

ing routers. Therefore, it is obvious that the total power consumption will be large if there

are many hops in the transmission path.

Throughput is another performance parameter, defined by the amount of data arriving

successfully at the destination in one unit of time. There are many ways to define it depending

on the specifics of the implementation. In general, the throughput T is defined as [65]:

T =(Total messages completed)× (Average message length)

(Number of IP blocks)× (Total time)(1.3)

Where total messages completed is the number of messages that have arrived at the desti-

nation node successfully; the average message length is in flits (in this report, we assume a

flit width of 20 bits); number of IP blocks refers to the number of routers that the message

passes by and the total time is the whole time the message spends in transmission, which is

equal to the time latency in 1.1. Thus, we see that if the number of IP blocks and total time

increase, the throughput will decrease.

4

Chapter 2

Problems in Traditional NoC

According to publications [25], [62], though an NoC has many important advantages,

it has a serious performance limitation due to the planar metal interconnects essentially re-

quiring multi-hop communication between any non-neighbor nodes. This causes high latency

and power consumption. In other words, the large number of functional nodes is the main

reason that limits the performance of a traditional NoC with wired communication. In the

future, the number of processors and memories in a single multi-core system will increase

to hundreds or even thousands [34]. Still using wired line as the communication channel

may cause a high latency, high power consumption and low throughput. To be specific, the

longer the path, more interference from other crossing paths will increase latency and power

to unreasonable amounts. In addition, the throughput will decrease. If there are hundreds

or thousands of nodes integrated on a single chip, the severity of the problem should become

evident.

5

Chapter 3

Some Solutions and Motivation for Studying Wireless NoC

In order to solve the problems outlined in the previous chapter, several solutions have

been proposed. In this chapter, I will give a brief introduction about the roadmap of the

development and some outstanding solutions for such problems. In addition, some represen-

tative examples will be discussed. However, the description here will serve as an overview

with detail of the technique appearing in the next chapter.

Some solutions, for examples, insert long-range direct links in a regular mesh network

by using conventional metal wires [61] to provide ultralow-latency and low-power express

channels between selected nodes [46], [48]. However, such solutions are still based on wired

communication channels. Although this seems to be a budding methodologies that signifi-

cantly improves performance when compared with an NoC system using the traditional wire

channel, it is not enough if the number of nodes on a chip increases to thousand or more

and the injection load is very high. According to the International Technology Roadmap for

Semiconductors (ITRS) [35] for the longer term, only improving in metal wire characteris-

tics will not satisfy the performance requirements and new interconnection paradigms are

needed [25].

Wireless offers an innovative solution. As mentioned, if we can decrease the number of

hops in a transmission path of a message, we may improve the performance in time latency,

power consumption and throughout. By using wireless, the transmission range can increase.

Therefore, for the same given transmission distance between two nodes, a message may go

through smaller numbers of wireless nodes.

Consequently, some approaches worth considering are 3D and photonic NoCs and NoC

architectures with multiband RF interconnect (RF-I) transmission lines [11], [69], [77], [99].

6

The basic idea of these alternative solutions is to insert some kind of express transmitting

channels to reduce the latency and power consumption. Although they can improve the

performance over that of any traditional NoC system without doubt, there is a lack of

effective technique to implement the hardware components. For example, the design of

transmitter, receiver, oscillator and filter and a high reliability integrated light source may

be needed. Although CMOS-based technologies can alleviate some manufacturing challenges,

they still need long physical lines to work as wave guides. Such unsolved problems are the

bottleneck in the use of on-chip wireless [12], [62]. So, though such emerging paradigms can

improve the performance, especially in latency and power consumption, of the traditional

wired NoC system to some extent, they are not matured and still need more study and

further research before the new solutions are deployed at a large scale.

That means, people should continue to look for other ways to realize the wireless com-

munication in NoC system. Here, I am going to introduce three leading approaches. Their

classification is based on the transmission method. They are ultra wide band (UWB) - based

WNoC, mm-wave-based WNoC and Carbon nanotube (CNT) - based WNoC system.

3.1 Ultra Wide Band (UWB) Based WNoC System

Different from other several developed RF interconnect technologies, UWB is a better

choice since it can achieve the low-power and low-cost implementation [98].

At the transmitter port, UWB-based interconnection uses Gaussian monocycle pulse

(GMP) generator to produce an ultra-short pulse so as to realize an extremely low power

spectral density (PSD). The modulation model used for the transmitted pulse can be pulse

position modulation (PPM) or biphase modulation (BPM). In addition, by using modula-

tion techniques for channelization and separation of users, for example, time-hopped PPM

(TH-PPM) and direct sequence coding, multiple access capability can be provided [71]. For

the received port, it is made up of a wideband low-noise amplifier (LNA), a correlator that

7

consist of a multiplier and integrator, an analog-to-digital converter (ADC) and synchro-

nization circuits. It should be mentioned that for the wireless NoC architecture, an ADC

including comparator and inverter buffer has been developed [74]. In this way, the perfor-

mance bottleneck for the receiver can be alleviated.

As a type of CMOS-based technology, the efficient integration of on-chip hardware com-

ponents is also a big challenge in UWB. To solve such problems, using higher frequency

instead of short-range wireless communication allows smaller antennas and increases to the

possibility of on-chip integration. As a result, the data rate can get 1.16Gbps for single

channel at a central frequency of 3.6 GHz [12], [24]. The data transmission protocol for an

UWB-based architecture is designed as a cross-layer. Such layers can accomplish most of the

functions of medium access control (MAC) network and transport layer in a traditional open

system interconnection (OSI) layered structure. Due to the characteristic of data transmis-

sion in WNoC, layers in WNoC should be designed specifically. For MAC layer, the key task

is solving channel contention and minimizing the collision probability. So, a synchronous and

distributed MAC (SD-MAC) protocol [85], [99], which is based on synchronized time frames,

has been raised to guarantee high efficiency, simplicity, robustness, fairness and quality of

service (QoS) capability. The transport layer does not consider the flow control and error

detection, so it just works as an interface between the network independent layer and the

network dependent layers. The main function of the network layer is to choose the data

transmitting path. In some recently studies [97], a simple location-based routing (LAR) has

been used. In this method, the path is only determined by the node’s current, its neighbors’

and destination’s locations, but not to have to maintain routing tables or network topology,

which saves power in some extent. In addition, to achieve 100% (guaranteed) delivery and

the QoS requirement in on-chip environment, a region-aided routing (RAR) protocol [96]

has been proposed.

8

The use of UWB has its own disadvantages. The achieved transmission range of UWB

based antenna is about 1mm [99], which means in a chip whose area is 20mm×20mm, we

will still need multi-hop communication when transmitting data.

3.2 A mm-Wave WNoC System

Another type of architecture used in WNoC uses millimeter-wave interconnections [18].

In order to achieve as much power gain for small area overhead as possible, a kind of zig-

zag antenna [52] is proposed. The data transmission process will also include modula-

tor/demodulator, serializer/de-serializer and amplifier. The modulation type used in this

architecture is on-off-key (OOK) modulation requiring the specific modulator and demodu-

lator [50]. Other hardware includes low noise amplifier (LNA), which consists of a two-stage

cascade amplifier with shunt-peaked load and an output buffer. The serializer/deserializer

(SERDES) devices are implemented with an oscillator block and multiplexer (MUX). Also

needed is a single pole double throw switch (SPDT), which can switch between the trans-

mitting and receiving modes of the transceiver.

For this type of WNoC system, the structure can be divided into two levels: several

nodes form a subnet and several such subnets will form the whole WNoC system. Each

subnet connects to another subnet through either wired or wireless communication. Within

a subnet, a node will connect to another node through wired interconnect. Therefore, the

whole WNoC system is a hierarchical architecture with two levels.

A shortcoming of this type of WNoC system is that due to the structure of the antenna

and the simple communication protocol, all communication channels work at the same fre-

quency. In a single-channel link there is only one wireless communication at a time to avoid

interference. In other word, though there are many wireless links in the WNoC system, only

one can be functional while others are in a rest state, which does not sound too efficient.

9

Table 3.1: WNoC system parameters.

UWB-based mm-based CNT-based

bandwidth 3.6GHz (central frequency) Tens of GHz around 500GHzdevices’ size millimeter order millimeter order micrometer order

transmission range not enough enough enoughnumber of channels multiple single multiple

3.3 CNT Based WNoC System

Carbon nanotubes (CNT) have proven to be a much better selection for use as antennas

for WNoC [40]. Some resent researches have shown that a single CNT can be used to make

several kinds of circuit components, such as antennas, modulator/demodulator [38], and

transmitter [39]. In [25], the CNT antenna has been used in a WNoC system.

CNT is developed by a chemical vapor deposition (CVD) method. In the fabrication

process, a localized heater is applied to avoid the preexisting CMOS layers being damaged

due to the high temperature CVD. The frequency range of CNT antenna can get to terrahertz

(THz) level, which has been investigated both theoretically and experimentally [40], [8]. If

the frequency range is in THz, the size of antenna will be small enough so that the area

overhead will be small. The bandwidth of CNT antenna can be high, unlike the mm-wave

antenna whose bandwidth is tens of GHz [31]. The bandwidth of CNT antenna is around

500GHz [25], which can afford a higher data rate.

What is more, there are three additional advantages of CNT-based antenna. First, it

can provide excellent directional properties. Second, since the skin effect in CNT can be

ignored even when the operating frequency is so high, the power dissipation of this type of

antenna is quite low [31]. Third, the multiband laser source can help the CNT antenna to

be assigned different frequencies in one communicating channel. The Table 3.1 compares the

UWB, mm and CNT based WNoC systems.

On the down side, a CNT-based WNoC system not readily deployable as it needs more

investigation. It though has apparent advantages than the other two options mentioned

10

above, especially its hardware overhead, which is significantly lower. So in the remainder of

the report, the description of WNoC system is mainly based on this type plus some additional

information based on other wireless types.

3.4 Other Needed Devices

Besides antennas, other devices are also indispensable for an integrated WNoC system.

For years of study and research and fast development of fabrication technologies, they may

proved to be practical.

The high-speed silicon integrated Mach-Zehnder optical modulator and demodulator

are commercially available currently [27], which can convert signals between electrical and

optical types.

For an oscillator, a kind of 324 GHz oscillator using 90 nm CMOS process [33] and a

410 GHz oscillator using 6M 45 nm CMOS process [75] have been used as on-chip wireless

interconnection transceiver. In [9], a low-power wideband wireless transceiver is designed.

The transmitter circuit consists of an up-conversion mixer and a power amplifier. The

receiver, which uses direct-conversion technology, is made up of a low noise amplifier, a

baseband amplifier and a down-conversion mixer. A voltage-controlled oscillator is used for

both transmitter and receiver.

11

Chapter 4

Technique Used in a WNoC System

This chapter introduce a general architecture of an WNoC system including the topol-

ogy/structure, wireless link insertion and transmission protocol. We provide a summary of

design topologies, communication protocols and routing schemes. Given that there are many

existing options, in this and in later chapters, I will describe some of the most popular ones

according to my recent study.

4.1 Structure and Topology

In WNoC system, the wireless link is a key component without any doubt. In general,

according to the combination of wired and wireless links in a system, the structure can be

divided into two types. One type is with pure wireless links for data transmission between

nodes and the other type contains both wired and wireless links. I should point out that, a

link whether wired or wireless mentioned here only refers to the data transmission channel.

Other connections, for the control channel and channel arbitration, are not considered.

4.1.1 Structure of a Pure Wireless NoC System

Such type of systems mainly depend on UWB transmission [100]. Every data channel

between nodes is wireless (not considering the control link since that does not work on data

transmission). Therefore, if the range of the wireless link is long enough, a node can connect

to any other node in the system by wireless directly, which means every data transmission

will be completed without intermediate hopping. In this condition, the system is formed as

a fully-connected network.

12

However, in a single frequency system (or because the wireless source has limited band-

width), a problem arises with the increase of transmission range: a fully-connected network

also implies each node will use more links, which will lead to greater channel arbitrations.

As a result more latency will be introduced. Additionally, it costs more area overhead for

more reserved wireless nodes. Therefore, a key point in the pure wireless WNoC system is

finding the best transmission range according to several factors, which include the topology,

number of nodes, injection load and routing algorithm.

4.1.2 Structure of Hybrid Wireless NoC System

Resent literature [9], [12], [25], [62] recommend hybrid WNoC more than the pure

WNoC. One reason is the limitation of the frequency band. Another major reason is the

efficiency and necessity for transmission between two short-distance nodes. Because the most

valuable characteristic of wireless communication is that it can decrease the number of hops

in a long range transmission, if there is none or just a few hops between two nodes, the

benefit of wireless is not so great over the wired transmission. Other reasons, such as the

device area overhead and device integration, are also significant. In a hybrid Wireless NoC

system, by designing a good combination of the wireless and wired links and suitable system

structure, one can alleviate the above problems.

The most popular structure of the hybrid wireless NoC is a kind of hierarchical structure.

The whole system is divided into several subnets, which can be considered as the bottom

level. Each subnet always contains less than 40 nodes (often the number of nodes in each

subnet will be 8, 16 or sometimes 32). The distance, which is treated as the number of hops,

between nodes within same subnet is short since a subnet can be called a “small-world”.

Typically, in the “small-world”, the average path length is no larger than logN , where N is

the number of nodes, and this make the topology so interesting for efficient communication

with minimum resources [7], [82]. Therefore, applying this feature, nodes in subnet are

connected by wires only, which is both efficient in communication and saved the wireless

13

resource. There is a hub in each subnet, whose function is to connect to other subnets

through both wireless and wired channels. The topology of the subnet can be 2-D mesh,

ring or star-ring, which can be 2 by 4, 4 by 4 or 4 by 8.

The top level is made up of all of the subnets. As mentioned above, communication at

the top level is wireless or wired and hubs in the subnets are the connection points. Due

to the limited resource of wireless which will be discussed later, the neighboring subnets are

connected by wire. The topology at this level is always ring, which means all subnets are

arranged in a circle and in every subnet nodes are arranged in 2-D mesh, ring or star-ring.

4.2 Wireless Link Insertion

In a chip, for a small world, though researchers can achieve a number of different fre-

quency bands to transmit data simultaneously, they still have not gotten enough frequency

bands for wireless channels when the number of nodes in a WNoC system is large. Moreover,

it is impossible to integrate so many antennas and other wireless interfaces when we consider

the device overhead and integration. Since the distance in subnet is constant once the source

node and destination node are determined, finding a way to allocate the limited number of

wireless link is a core problem in enhancing the performance.

The key task is minimizing the average distance between nodes. One of the ways is

using simulated annealing (SA) [45] to obtain an optimal configuration from placements of

the wireless links and hubs [9], [25], [62]. By using SA, designers get the most optimal pairs

of source hub and destination hub in wireless while other pairs are still wired. In this way,

selecting an optimal configuration for a better performance.

4.3 Routing and Communication Protocol

We will study this in two parts, one is the protocol in pure wireless NoC system and

the other is the hybrid wireless system. In either case, since there are so many proposed

14

solutions, considering the key point to limit the length of this report, we will introduce some

basic and typical proposed methods.

4.3.1 Protocols in Pure Wireless NoC System

LBR Routing Scheme

For RF nodes, the transmission range is determined. A routing algorithm named

Location-Based Routing (LBR) has been proposed [100]. It is a type of static routing

scheme. The basic idea of this algorithm is that in the possible transmitting range, a source

node sends data to the node that is nearest to the destination at every step. A process to

improve the routing efficiency partitions the whole net into four parts, namely, quadrants,

depending on the positions of both source node and destination node. By calculating the

differences of X coordinates and Y coordinates of source node and destination node, we de-

termine which quadrant the destination is in. Next, we find which neighbor in this quadrant

will bring us closest to the destination.

X-Y Routing Scheme

This algorithm is an adaptation of pure XY-routing. In this scheme, source node trans-

mits data in X direction at first. When the current node’s X coordinate equals that of the

source, data is sent in Y direction until it reaches the destination node. In both X and Y

routing, node transmits data to the one nearest to the destination.

In this structure, a node may be both a sender and receiver simultaneously, and may

intend to receive from and send to multiple channels at the same time as well. Moreover, a

sender/receiver can send/receive on one channel. Therefore, in order to alleviate the potential

collision in both of the conditions, an arbitration scheme is applied. In this scheme, each

node maintains a request arbitrator and an authorization arbitrator. Considering the point

of view of the receiver, the request arbitrator is used to decide upon a sender to receive data

from, if there are several senders wanting to send data to the same receiver; once the winning

15

sender is found, the authorization arbitrator sends acknowledgment back to the winner.

In the meantime, the winner may also have to send data to others in multiple channels

and may have received several authorizations from different receivers. Then, authorization

signals will be used by the sender and only one will be accepted. Finally, the data is sent

out. Additionally, to simplify the requesting arbitration, researchers apply a receiver-based

coding scheme [80].

SD-MAC Protocol

Another proposed protocol is the synchronous and distributed medium access control

(SD-MAC) protocol [99], which is used on MAC layer. In general, the principle of this

protocol is similar to that of the former one. It consists of control and data transmission

parts. One difference, however, is that the arbitration section, which is called competition

in SD-MAC, is designed to be more specific. The competition is executed by a competition

frame, which is further divided into three parts in the SD-MAC protocol.

The first part is called the initialization period (INIP). The function of this period is

to record the information of each node and its neighbors. It has two segments. The first

one records the four states of the current node: transmitter (TX), receiver (RX), inactive

node (InA) and not determined (SND). The state is recorded in a 2-bit register. The second

segment keeps the information about the current node’s potential senders (PSR) and survived

senders (SSR). The potential senders are nodes intending to send data to the current node,

whereas the survived sender is the winner among the potential senders. This segment is

represented by an N -bit register, where N is the number of potential senders.

Following the INIP is the contention period (CP). This period works on the channel

contention process, which mainly focuses on providing the necessary data to decide upon the

survived sender. In other word, makes preparation for the channel access. In this interval,

senders generate an N -bit random number to help make the decision; and by calculating

16

bitwise AND of RX and SSR and updating the SSR register round by round, one gets the

final winner - the only survived sender.

The last part of the competition frame is channel access authorization period (CAAP).

This period truly is the time to process the decision described above. When the CAA period

finishes, the link is built. The SD-MAC protocol can avoid collision very well. Additionally,

it can solve the exposed terminal problem when data is transmitted in parallel [99].

4.3.2 Protocols in a Hybrid Wireless NoC System

Since there are both wired (within the subnet or between subnets) and wireless (between

subnets) links in a hybrid wireless NoC system, the protocol and routing algorithm usually

depends on these two types of communication.

Routing Scheme Based on Comparison of the Path Length

A simple routing protocol described in [12] is based on comparing the numbers of hops

between the wireless and wired links.

If the two nodes are in the same subnet, they should connect through wire only. If two

nodes are in different subnets, the source node will select a method according to a signal sent

via wire from the wireless router in the subnet where the source node is located. Apparently,

the purpose of the signal is to notify whether the intending wireless link is busy or idle. If

the wireless link is busy, data should be sent by wire; if it is not, we proceed to the next step.

The second step targets on balancing the usage of wired and wireless links. The traveling

distance of both wired and wireless links is counted in number of hops. Before choosing

the communication method, source node calculates the distances through wireless and wired

links by using the destination ID and uses the smaller of the two.

Obviously, when nodes are connected only by wire, the data will traverse the wired

links; if nodes use wireless links, source node should still link to the wireless router in the

same subnet via wire at first, and then transmit data by using wireless.

17

Adopting a Routing Scheme

In [9], a method for adopting a routing strategy is presented for a mm-based WNoC

system. For transmission in the subnet, two types of topologies are discussed, mesh and

star-ring. In mesh subnet, a deadlock-free dimension order routing is used. In the star-ring

topology, two sub-conditions are discussed. If the distance, say, the number of hops, is less

than two, the routing path is along the ring. If data must travel more than two hops, the

flit should go through the center node so that the total distance can decrease to two. In

order to avoid deadlock in this kind of topology, virtual channel management scheme from

Red Rover algorithm [20] is adopted. In this algorithm, the whole ring is divided into two

equal sets of nodes and each set has its own virtual channels. For communication at the

top level, say, between subnets, the condition is classified by the usage of wireless interface

(WI). If both the subnets that source node and destination node are in have WI, this pair

will connect in wireless link. If only one of them has WI whereas the other one does not, the

data will be sent to the nearest subnet that has a WI and then it will go to the destination

(for condition that source subnet does not have WI but destination subnet has), or data will

be sent to the WI subnet which is nearest to destination subnet and then to the destination

on wire (for condition that destination subnet does not have WI but source subnet has). If

neither of the source nor the destination have WI, routing path will be chosen as the smaller

one among the wireless link and wired link. However, in this situation, the WI will be hot

spots. Therefore, a token flow control [47] is used to resolve this problem. Basically, tokens

reflect the states of the buffers of each input of a WI node. Every input has a token and

if the token is greater than a fixed threshold, then the token will turn on and indicate that

this input is available, whereas the token will turn off if it is smaller than the threshold and

tell the sender that the corresponding input is busy. The routing strategy used at this level

is also classified as two kinds: for hubs that do not have WIs, a dimension order routing is

adopted; for hubs that have WIs, South-East routing algorithm, which has been proven to

be deadlock-free [61] is used.

18

In this architecture, all wireless communications are in the same channel. To avoid

the interference and alleviate the channel contention, an arbitration mechanism is designed

to guarantee that the source node can reach its true destination. A token strategy is also

proposed here. The difference between this token method and the one described above is

that the wireless token can broadcast data in flits into the wireless medium so that all other

WI antennas will receive the signal. The one that matches the destination address will use

the channel. Then, the wireless token will go to the next hub that needs arbitration.

Multi-Channel Protocol and Routing Scheme

In [99], different frequency bands can be assigned by using multiband laser source that

excites CNT antennas. Thus, an opportunity to use frequency division multiplexing (FDM)

to create the dedicated channels between source and destination node comes up. It is realized

by using CNTs of different lengths. Additionally, some researchers [40], [32] have demon-

strated that this kind of CNT has a high directional property, which is very suitable for using

in directed channels. The number of different frequencies available for dedicated channels

has been reported as 24 [49], which means 24 wireless channels can be used simultane-

ously in a single WNoC system without any frequency interference and channel contention.

Besides FDM, time division multiplexing (TDM) is adopted in each frequency band. To

sum up, in this structure, a wireless antenna uses one of the twenty-four frequencies, and

in the dedicated frequency band, the antenna sends data, whose destinations are different,

by using several time slots. The modulation scheme in this system is non-coherent on-off

key (OOK), which leaves the system with no necessity to have complex clock recovery and

synchronization circuits.

In a specific case reported in [99], for a top level with ring topology, a flit, the unit

of transmitted data, is assumed to contain 32 bits and the number of time slots is 8 per

dedicated channel. Therefore, each time slot contains 4 bits. With the optical modulator

proposed in [27], 10 Gbps data rate can be provided per channel. Therefore, in this case, the

19

length of each time slot is 0.1 ns, corresponding to the 10 Gbps. The number of frequency

bands is 4. That means, in the specific WNoC system, the bits in each time slot can be

transmitted over four channels.

For each transmission, the path is predetermined at the source hub so that there is no

possibility of deadlock. The information of the path, such as the address of the intermediate

hub, is kept in the header flit, which is followed by the remaining data flits. When sending

the data, it is just likes a “worm” along the path. Since the wireless link can be treated as

a short cut when compared to the wired link, a flit may chose more than one wireless links,

initially. Though in this system, there are several simultaneous wireless channels, it is still

hardly comparable in size to the functional nodes. In order to alleviate hot spots and get

a best trade-off between router complexity and network performance [99], only a path with

one wireless link will be selected as the true transmission path. When there are several paths

with the same number of hops, the one with the wireless link is chosen, considering the low

power consumption of wireless communication.

Another routing mechanism that contrasts the above one can also be incorporated with

the combination of FDM and TDM scheme. For this routing approach, the path is not

determined at the source node in once. Instead, at each current node, an evaluation is done

to select a possible next best step.

20

Chapter 5

A Simulator for NoC System

Currently, there is no existing general purpose simulator for NoC systems. Most simu-

lators are designed and used by users for their specific purpose. These simulators are always

cycle accurate. Some of them are developed based on SystemC.

SystemC is maintained and developed by OSCI (Open SystemC Initiative). It is a kind

of system-level description language that is based on C++. SystemC is built on a series of

C++ class structures. In fact, it is a library in C++. SystemC is highly suitable for the

system-level simulation.

There are also some open source simulators on the Internet. One of them is named

Noxim [60], which is developed by University of Catania, Italy. This simulator is written

using C++ language, which is mainly based on SystemC. By using Noxim, users can get

some basic parameters of a NoC system, such as throughput, delay and power consumption.

Noxim should be run in Linux environment. However, after some discussion between the

author and others, we found some errors during the compilation if the version of g++ (a

compiler of C++) is beyond g++ 4.0. The reason may be that there is a conflict between

SystemC and g++, which is a bug of SystemC.

Nirgam is another cycle accurate simulator for NoC, which is also based on SystemC.

By confirming topology, switching mechanism, buffer, clock rate, routing algorithm and

other parameters that depend on the objective of simulation, user can get the latency and

throughput of an NoC system. Nirgam also runs in the Linux environment.

The above mentioned simulators focus on traditional NoC systems. To simulate a WNoC

system, some necessary changes should be made since there are differences between the

WNoC and NoC architectures, such as the channel’s character, number of nodes, combination

21

of protocols (intra- and inter- subnet) used in the system and routing algorithms. For

example, WNoC platform used in [12] contains a 100-core system. The system is further

divided into subnets and each of them is a 5×5 mesh network. To connect pairs of neighboring

subnets, the author used the high-speed wireless link, which is different from the wired link

in channel character. In each subnet, 24 wired routers are employed, which will use different

routing algorithm and protocol from a traditional NoC system. Still, there is no general

simulator for WNoC for public use. Most organizations use the simulators designed by

themselves for a specific purpose or task.

In this work, I use Booksim interconnection network simulator. Booksim, a cycle-

accurate interconnection network simulator, is developed by Stanford University. The current

major release, Booksim 2.0, supports a wide range of topologies such as mesh, torus and

flattened butterfly networks, provides diverse routing algorithms and includes numerous

options for customizing the network’s router micro architecture [5]. Similar to most NoC

simulators, Booksim is written in C++ and should be run under Linux. However, for windows

user, it can also work with the newest version of Cygwin [15]. To download the simulator, we

should use subversion (SVN) repository to get the source from the web. After the download,

typing “make” to build the simulator at the “/src” directory. When using Booksim to

simulate performance at different conditions (injection rate, topology, size, etc.), we can

just set the corresponding values in the configuration file and then execute the modified

file. When Booksim is running, the process can be divided into three stages: warm up,

measurement and drain. In practice, the initial state of a network system cannot be ideal,

which means no data is transmitted in the network and all nodes are free. On the other

hand, there must be some functional nodes in the whole system. Therefore, in order to begin

the simulation as what it would be like in practice, the system should be injected with some

data and left running before the simulation experiment starts. That is the function of the

“warm up” stage. In the “measurement” stage, the simulator is doing the simulation. Once

the “measurement” stage is completed, all the measurement packet are drained from the

22

network before the results (latency and throughput) are printed. That is the “drain” stage.

The details about how to use Booksim will be introduced in Chapter 7.

23

Chapter 6

Factors and Parameters

As mentioned in the previous chapters, the performance of an NoC/WNoC system

mainly contains the throughput, latency and energy consumption. There are many factors

that affecting such parameters.

Factor topology determines the connecting form of the system. Additionally, the size,

or in other words, the number of nodes, can be set in the topology factor. For mesh or

torus topologies, we always use an N × N network. To be more specific, we can also set

the number of nodes that share a single router in the topology. Last but not least, for some

kinds of topologies, such as mesh and torus, there is even a probability of link failure, which

can also affect the performance.

Injection rate is the rate at which packets are injected into the simulator. In other words,

injection rate is added to tell the simulator how many packets to inject per simulation cycle

per nodes on an average. If the rate is too high, which means in practice there is a high

injection of data into the WNoC/NoC system, the latency will become huge.

Flow control is another factor that should be considered. It refers to the number of

virtual channels [16] per physical channel and the depth of each virtual channel; the unit is

flit. Such factors can relate to the latency and throughput of a system. Routing algorithm

is another key aspect. For instance, if the router itself is deadlock-free, which means it can

use the virtual channels freely, the latency will be smaller; if we have to partition the virtual

channel to avoid the deadlock, there must be some additional latency introduced.

24

Chapter 7

Experimental Processes, Results and Analysis

This experiment contains both wired and wireless NoC systems. Based on the results,

some analysis and estimation of WNoC will be given. Besides considering the size as the

main factor, this experiment mainly focuses on the relationship between the flow control

and the performance, say, latency, so that the results will point out which technical elements

may contribute to it. At last, a comparison between performances of traditional and wireless

NoC systems plus some discussion will be given.

7.1 Experimental Process

Booksim, an interconnection network simulator, is used in this project as mentioned

before. Once download and compile it successful, the user can find that there are some

important files in the /booksim/trunk/src directory. Folder named “examples” contains

several configuration files, which contain data on different topologies that can be simu-

lated. In each configuration file, there are some parameters should be set as they will affect

the simulation result. By using those files and changing or resetting parameters in them,

the user can simulate the performance of a variety of NoC systems. Another useful file

is “booksim config.cpp”. It specifies default values in case the configuration file does not

specify any values.

Before we run the simulation, we may set some parameters of the network, such as the

“injection rate”. By typing “vi [configfile]” in “/src/examples” directory, we can enter

the file to set different parameters that we need. “[configfile]” is the name of the file that

contains the configuration information for the simulator. For example, the command “vi

mesh88 lat” means that the user can set parameters in the “mesh88 lat” file, as shown in

25

Figure 7.1: Example configuration file for simulating a mesh NoC system.

Figure 7.1. As this figure shows, in the configuration file there are several parameters, such

as “num vcs”, “vc buf size” and “injection rate”. All of these parameters can be changed

depending on the specific task. To change the parameters that do not appear in configuration

files, the user can set it in the default file, named “booksim config.cpp”, as was mentioned

previously. If the user changes parameters in the configuration file, it is fine to just save

and quit the file; if the user sets a new value in the default file, he/she should remake the

simulator so that it is guaranteed that the new value has been stored. Here, some important

parameters meanings are defined: “num vcs” means the number of the virtual channels per

router; “vc buf size” means the size of buffer, the depth is in flit; “traffic” defines the

traffic patterns, the option “uniform” means each source sends an equal amount of data

to destination; for injection rate, its unit can be either flit/cycle/node or bit/cycle/node.

Each flit contains 20 bits. In this report, we use the former definition. For example, if the

injection rate is 0.15, it means that during each cycle, 0.15 bit is injected in every node.

26

Figure 7.2: Simulator output from running the examples/mesh88 lat configuration file.

Once the parameters have been set, we can begin to simulate. In order to start it,

type command “./booksim [configfile]” in the “/src” directory. Let us take “mesh88 lat”

as an example. The outcome is shown in Figure 7.2. At the bottom of the figure, the

program prints the “average latency”, “average accepted rate”, which means the average

throughput, “min accepted rate” and “average hops”. The unit of latency is cycle; the unit

of throughput is flit/cycle/node.

Such configuration can generate a single data point. According to Figure 7.2, it can

be a point of “injection rate vs. latency” line. For this point, it indicates that when the

injection rate is 0.3 bit per cycle per node, the average latency is about 40 cycles. Then, by

running the simulation for many increments of “injection rate”, an average latency graph

can be obtained.

The next section will examine relationships among injection rate, latency and number

of virtual channels for different topologies. Additionally, a discussion about the trade-off

between number of nodes per router and the distance between two nodes is given. At the

end of the section, an analysis of the performance of WNoC system is also presented.

27

Figure 7.3: A 4× 4 mesh topology.

7.2 Experimental Result and Analysis

For all of the following simulation on latency, I uniformly used some basic conditions:

the size of each virtual channel’s buffer is 8 flits, and traffic pattern is uniform. The routing

algorithm is dimension-ordered. Other parameters, such as speed-up in input and output,

delays in each stage and type of allocator, all have default values.

7.2.1 Latency vs. Injection Rate

This part shows the results of simulation of the latency curves for different injection rate

levels. The first topology is mesh, whose structure is illustrated in Figure 7.3. This figure

shows a 4× 4 regular mesh topology, the squares represent the nodes, including caches and

processors; the black points represent routers, which are connected by wired line. From the

figure, we can easily find that each node has its own individual router.

When we simulate the performance by changing the values of injection rate and repeat

the simulation for each case, we can get a series of data points. Connecting them together,

a latency graph is obtained as shown in Figure 7.4 for different numbers of virtual channels

(4, 8 and 16). Each latency curve represents the performance for a specific number of virtual

channels. There are 64 nodes in the system, which make up an 8× 8 regular mesh topology.

28

0.2 0.22 0.24 0.26 0.28 0.3 0.32 0.34 0.36 0.38 0.4 0.42 0.4450

100

150

200

250

300

350

400

450

500

550

injection rate (bit/cycle)

late

ncy

(cyc

le)

Number of virtual channel: 4Number of virtual channel: 8Number of virtual channel: 16

Figure 7.4: Latency vs. injection rate in mesh topology.

From the simulation result, we observe that, the latency increases as the injection rate

goes up. The latency starts from about 60 cycles when injection rate is 0.2 bit per cycle. In

general, more virtual channels each router has, lower is the latency in the system. Therefore,

an existing rule is evident here, that is, for a regular mesh topology the utilization of virtual

channel can help the system to reduce the latency. However, if the injection rate is below

0.32 bit per cycle, the number of virtual channels will not affect the latency, or at most,

affects it only mildly. Once the injection rate exceeds 0.36 bit per cycle, the difference in

latency due to the number of virtual channels becomes more obvious. What is more, for

all three curves, 0.36 is the “take off” point. Which means, that for an 8 × 8 regular mesh

topology, when the injection rate exceed 0.36 bit per cycle, the contribution of injection

rate becomes obvious in the form of rapidly increasing system latency. However, there is no

denying that the more virtual channel each router in the system has, the flatter the cure

becomes. In addition, when each router has 4 virtual channels and the injection rate goes

29

Figure 7.5: A 4× 4 mesh topology with concentration 4.

beyond 0.4 bit per cycle, the system’s latency becomes huge and we may treat it as infinite

or unusable. Thus, in Figure 7.4, the blue line stops when the injection rate equals 0.4. For

systems that have routers each with 8 × 16 virtual channels, the maximum injection rates

are set to 0.44. Beyond that level, the latency will be out of control.

A kind of variant of the regular mesh topology is named concentrated mesh topology

(called “cmesh” for simplicity in this report). A different from of regular mesh in which each

node has its own routers, while in concentrated mesh, several nodes share a single router.

The term “concentrated” refers to how many nodes share a single router. For examples,

if we say a system has concentration 4, that means, in the system, 4 nodes share a single

router. The topology of cmesh with concentration 4 is shown in Figure 7.5.

Here, we find that each router is shared by four nodes so that the maximum distance

between any two nodes is 2 (count in hops). Comparing with the regular mesh that contains

the same number of nodes and whose maximum distance is 6 as can be verified from Fig-

ure 7.3, we find that the distance has dropped, significantly. I used Booksim to simulate the

average hops in both these topologies. Figures 7.6 and 7.7 give simulation results for 8× 8

mesh and cmesh. For these two topologies, the longest distances are 14 and 6, respectively.

From the figures, as we can see that in regular 8× 8 mesh, the average hops exceeds 6 while

it is around 3 in concentrated mesh (cmesh) topology of the same size.

30

Figure 7.6: Average hops in 8× 8 mesh.

Figure 7.7: Average hops in 8× 8 mesh with concentration 4.

Having seen the difference in average hops for concentrated mesh and regular mesh

topologies by running the “cmeshconfig” file, we then examined the performance of the

concentrated mesh. This is shown in Figure 7.8, Once again, the figure is a latency curve

for different values of injection rates. Different lines stand for different sizes of the system.

In each system, every router has 16 virtual channels. The general trend is the same as for

the regular mesh topology. Higher injection rate leads to higher latency. When the injection

rate is small, say, not larger than 0.11 bit per cycle, the latencies in systems of different sizes

are almost same, i.e., around 25 cycles. For different sizes, the bounds of injection rate are

different: in a 64-node system, the maximum injection rate is 0.26 bit per cycle, in 144-node

system, it is 0.17 bit per cycle; while in 256-node system, the maximum accepted injection

rate is only 0.13 bit per cycle. For a specific injection rate, the 64-node system has the

smallest latency while the 256-node system has the largest latency.

The preceding results allow us to verify another existing rule, that is, in the same

topology, the larger the size of the system, larger is latency for the same injection rate.

However, although these three systems have different injection rate bounds, their latency

“take off” points are almost the same, which are all at about 50 cycles. Before that level,

the latencies increase slowly or barely increase; once the latencies exceed 50 cycles, they go

up dramatically. That says that in a cmesh network, until the injection rate stays below a

specific level, the latency will increase very little in spite of increase in the injection rate.

31

0.1 0.110.120.130.140.150.160.170.180.19 0.2 0.210.220.230.240.250.260

50

100

150

200

250

300

350

400

450

injection rate (bit/cycle)

late

ncy

(cyc

le)

Number of nodes: 64Number of nodes: 144Number of nodes: 256

Figure 7.8: Latency vs. injection rate in cmesh topology.

As the source of the latency is the injection rate of the system, distance between nodes

and the system’s concentration are two other factors that lead to increased latency. So, it

is natural to ask, which factor will affect a mesh system the most; or what are the best

trade-off values? In order to find an answer, further comparison between regular mesh and

concentrated mesh was done. The sizes for both were 64 nodes and each router had 16

virtual channels. In the concentrated mesh system, every router is shared by 4 nodes. The

result is shown in Figure 7.9.

In Figure 7.9, the regular mesh curve is flatter than the concentrated mesh curve. Before

injection rate equals 0.24 bit per cycle, the regular mesh has larger latency than concentrated

mesh for every injection rate. Once the injection rate surpasses 0.24, the concentrated mesh

curve starts to drown out the regular mesh curve. Therefore, although in the beginning

regular mesh has a larger latency, its maximum bound of injection rate is much larger than

the concentrated mesh. To sum up, when the injection rate is low, it is better to use a

concentrated mesh; if it is in a heavy load system, a regular mesh system is wiser choice.

32

0.2 0.22 0.24 0.26 0.28 0.3 0.32 0.34 0.36 0.38 0.4 0.42 0.440

50

100

150

200

250

300

350

400

450

injection rate (bit/cycle)

late

ncy

(cyc

le)

nodes per router: 4nodes per router: 1

Figure 7.9: Comparison between regular mesh and concentrated mesh.

No matter how we improve the structure of a traditional NoC system, the performance

seems not so ideal. Therefore, next, a simulation of wireless NoC System is done under a flat-

tened butterfly topology. The flattened butterfly topology is a kind of cost-efficient topology

[43]. The structure of the topology and its router placement are shown in Figure 7.10 [42].

For a 64-node system, each router is shared by 4 nodes. In Figure 7.10(a), each cycle

stands for one node. Four of them are in one group and connect to the same router, which

is represented by a square. For each router, besides connecting to the 4 nodes, it also has

6 more ports. Three of them connect to the other 3 routers in the same dimension, say x

dimension; the other 3 connect to routers in dimension 2, that is y dimension. The router

layout is shown in Figure 7.10(b). Thus, no matter whether routers are in rows or in columns,

all of them are fully connected. Although it is an 8 × 8 topology, the maximum distance

between any two nodes is no larger than 2. That makes it closer to a wireless environment in

several ways. First, the flattened butterfly topology is also a kind of concentrated topology

33

Figure 7.10: (a) Flattened butterfly topology consisting of 64 nodes; (b) corresponding routerlayout.

(the concentration is 4), that is like in a WNoC system, every wireless interface is shared by

several nodes in the same subnet. Second, the maximum distance in flattened butterfly is 2,

which works on reducing the distance between nodes as the WNoC does. Therefore, I did a

wireless simulation by using flattened butterfly network.

Figure 7.11 shows the latency curves for different injection rates in different system size:

one system consists of 64 nodes while the other system has 256 nodes. Still, a comparison of

different numbers of virtual channels is included in these graphs. There are some similarities

the two systems. First of all, since the topology in WNoC is also a kind of concentrated

topology (in this experiment, the concentration is 4), shapes of curves in both graphs are

similar as they are in Figure 7.8. Even the places where the “take off” points appear are

34

0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.90

100

200

300

400

500

600

injection rate (bit/cycle)

late

ncy

(cyc

le)

Number of virtual channel: 4Number of virtual channel: 8Number of virtual channel: 16

(a)

0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.90

100

200

300

400

500

600

injection rate (bit/cycle)

late

ncy

(cyc

le)

number of virtual channels: 4number of virtual channels: 8number of virtual channels: 16

(b)

Figure 7.11: Latency vs. injection rate in flattened butterfly topology consisting of (a) 64nodes and (b) 256 nodes.

still similar to that in concentrated mesh topology, which is at the level that latency equals

around 50 cycles. Secondly, a big improvement of the WNoC system is the maximum

accepted injection rate. In these systems, the injection rates can get to 0.9 and 0.85 bit

per cycle, respectively. Since the injection rate also has a significant impact on throughput,

WNoC system will have much higher throughput when compare with traditional NoC. The

tiny difference on the maximum injection rate is caused by the different sizes of these two

systems. This advantage is due to the reduction of the distance between any two nodes

when comparison with other topologies, that is, the largest distance between any nodes is

35

no bigger than 2 in this kind of topology. Therefore, we conclude: if the number of nodes in

a transmission path is smaller, fewer routers will be used and hence smaller latency will be

introduced. Last but not least, when the injection rate remains below some upper bound,

the number of virtual channels will not have so much effect on the latency. In a system

consisting 64 nodes, before injection rate reaches 0.7 bit per cycle, the latencies are almost

the same no mater how many virtual channels a router has. In a system consisting 256

nodes, the difference happens when the injection rate is 0.5.

However, there is still a difference between these two figures, it is the effects of the

virtual channel on the injection rate’s upper bounds. In system consisting 64 nodes, we can

find that although the three curves reach different stop points, they do not leave far away

from each other: the maximum injection rate is 0.8, 0.85 and 0.9 bit per cycle for 4, 8, and

16 virtual channels, respectively. However, in system consisting 256 nodes, there is a big

difference between 4 and 8 virtual channels. Such phenomenon indicates that the virtual

channel will make more contribution on increasing the maximum injection rate value in a

larger size system. For the above specific examples, the gap happens when virtual channel’s

amount increase from 4 to 8 in a 256-nodes system. Furthermore, in both figures, the shape

of 8-virtual channels curve likes the shape of 16-virtual channels curve very much. This

result says, when the amount of virtual channels surpass 8, the effect on improve maximum

injection rate will not be so obvious.

From the above comparison we conclude that, considering their latencies and the through-

puts, a WNoC system will provide a big improvement when compared to traditional wired

NoC system.

7.2.2 Latency vs. Virtual Channels

A careful examination of data in the previous section will show that virtual channels

also contribute to latency under certain conditions. Now we examine this issue in detail.

Figure 7.12 shows simulated latency as a function of the number of virtual channels in regular

36

2 3 4 5 6 7 8 9 10 11 12 13 14 15 1650

100

150

200

250

300

350

400

450

number of virtual channels

late

ncy

(cyc

le)

injection rate: 0.2injection rate: 0.25injection rate: 0.3injection rate: 0.35injection rate: 0.4

(a)

2 3 4 5 6 7 8 9 10 11 12 13 14 15 160

50

100

150

200

250

300

350

400

450

500

number of virtual channels

late

ncy

(cyc

le)

injection rate: 0.4injection rate: 0.5injection rate: 0.6injection rate: 0.7injection rate: 0.8

(b)

Figure 7.12: Simulated latency vs. number of virtual channels for (a) regular mesh and (b)flattened butterfly topologies.

mesh and flattened butterfly topologies. The latter can be considered similar to a wireless-

like environment. Both systems have 64 nodes. For either topology, the injection rate is

varied from some maximum value to a value below which the shape of the latency curve

does not change.

An examination of Figure 7.12 reveals that for high injection rates, virtual channels can

bring the latency down. In both topologies, when the injection rate drops to about half of its

maximum value, i.e., 0.25 bit per cycle in regular mesh and 0.5 bit per cycle in wireless-like

topology, increasing the number of virtual channels does not contribute much.

37

Next, we examine differences between the two topologies. First, the latency for wireless-

like topology shows a more monotonic decrease. But for regular mesh, we see monotonic

decrease with fluctuations, especially when the injection rate is high. Another difference

is that, in regular mesh network when the number of virtual channels increases the best

achievable latency rises with injection rate. However, in the wireless-like network, the best

achievable latency is more or less independent of the injection rate (except for some very

high injection rate) once the number of virtual channels exceeds 4 or 5.

To sum up, when a system has high levels of injection rate, virtual channel will play a

more important role. What is more, in different systems the benefits vary. In this simulated

experiment, virtual channels are found to be more beneficial in a wireless-like network than

in a regular mesh network.

7.2.3 An Analysis of WNoC System

As mentioned above, I treated the flattened butterfly topology as a representative model

for a wireless-like network. Therefore, although in this experiment all simulations use wired

environment, we can draw some reasonable inferences applicable to wireless NoC systems.

Recall that in flattened butterfly topology the maximum distance is 2, similar to a link with

two directly-connected nodes in a large network. Further, in WnoC every wireless router

is shared by several, often more than 10, nodes. That is just like a traditional NoC with

concentration, though the router is different from the one used in wired NoC system. Indeed,

there are some differences. For examples, just to name a few communication protocols and

bandwidths differ.

Figure 7.13 shows the simulation of a 64-node system with 4, 8 and 16 virtual channels,

respectively. No matter how many virtual channels are used, the WNoC-like system performs

much better than the traditional NoC system. We observe that at the same injection rate,

the latency of a traditional NoC (blue curve, regular mesh) is much higher than that of

a WNoC-like system (red curve, flattened butterfly). The maximum permissible injection

38

0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.90

50

100

150

200

250

300

350

400

450

500

injection rate (bit/cycle)

late

ncy

(cyc

le)

regular meshflattened butterfly

0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.90

100

200

300

400

500

600

injection rate (bit/cycle)

late

ncy

(cyc

le)

regular meshflattened butterfly

(a) (b)

0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.90

100

200

300

400

500

600

injection rate (bit/cycle)

late

ncy

(cyc

le)

regular meshflattened butterfly

(c)

Figure 7.13: Comparison between regular mesh and flattened butterfly with (a) 4 virtualchannels, (b) 8 virtual channels and (c) 16 virtual channels.

rate for a traditional NoC is much smaller than that for a WNoC-like system. Evidently,

a WNoC system can process much more data in the same time. Additionally, in a WNoC-

like system, the behavior at the “take-off” point is sharper. That means that, until some

maximum injection rate is reached, any increase in injection rate will almost not effect the

latency. On the other hand, a traditional NoC system displays a continuously rising latency

with increasing injection rate. In a real operational environment where the activity may be

a function of time, these characteristics will produce a higher and more stable throughput

behavior for the WNoC system.

39

Chapter 8

Conclusion and Future Work

As a promising and burgeoning technique, NoC systems have been a hot topic of study by

many researchers and organizations in recent times. Such researches cover from the system-

level down to the device-level and numerous achievements have been reported. Considering

the problems of multi-hop communication and improve the performance of a large NoC

system, the idea of inserting wireless links offers a potential solution. Therefore, our study

of wireless network-on-chip (WNoC) systems is timely.

8.1 Work Completed in This Report

The present report is a survey of the field. It provides a large list of references. In addi-

tion, simulation experiments and conclusions, some of which may be original, are reported.

For the lack of a simulator for WNoC, we have simulated a wired NoC with butterfly topol-

ogy that has an upper bound on the communication distance between nodes. We believe

our conclusions will prove to be valid for WNoC.

8.1.1 Device-Level

For the device-level, the research on the basic components for NoC or WNoC systems be-

gan years ago and progressed through immense advancements of the CMOS technology. For

example, the antenna, one of the most important components in a WNoC system, whether it

is a mm-wave antenna or a CNT-based antenna or any other type of RF antenna, is feasible

today. What is more, beside those, many other kinds of antenna are also be developed [21],

[30], [70], [102]. The main purpose and contribution of such papers is to improve the per-

formance of the antenna. Another key component of NoC and WNoC that have been a

40

focus of research is the router [10], [17], [54], [72], [73], [87], [88]. These researches consider

router’s performance and other characteristics such as power consumption, throughput, la-

tency and fault tolerance. The also study routers with various topologies including 3-D.

Nowadays, the wormhole router is popular in NoC or WNoC systems since it can increase

the throughput of the whole system due to its architecture. Last but not least, the utiliza-

tion of virtual channels also makes a significant contribution to the development of NoC and

WNoC systems. By using the virtual channels, the performances in throughput and latency

can have a huge improvement. Other communications devices have also been developed and

include transceiver circuit [9], [83], [90], which contains transmitter, receiver, modulator and

demodulator. There are also detailed designs available for the switch [65].

8.1.2 System-Level

At the system level, types of topologies, communication protocols, quality of service

(QoS) of the communication, routing schemes and wireless insertion have been worked out

at various levels of detail.

Topology

For WNoC systems, most research has focused on commonly known topologies, such

as 2-D mesh, 3-D mesh, SPIN, folded torus, fat tree, ring and star-ring. Reference [26]

gives a brief review of some of those architectures. The introduction mainly focuses on the

scalability and structure definition. The most popular topology used in research is 2-D mesh

for wired NoC systems. For WNoC systems, 2-D mesh is used for wired parts (commonly

known as subnets) and ring/star-ring is used for wireless parts.

Of course, there are other topologies for an NoC system. References [6], [14] and [42]

study the tree topology (fat tree and flatted butterfly tree), focusing on throughout, latency

and energy efficiency. Deadlock recovery and low power consumption with torus topology

41

is proposed in [78]. The Clos interconnection network is studied in [94], and this paper

contributes to improvements in the performance of NoC systems.

Beside those common topologies, other topologies are proposed and studied for specific

purposes [1], [36], [41], [79], [84], [93]. The purpose can be throughput, mapping, routing

algorithm, power consumption or link optimization. Although these topologies are not so

common as those that focus on the whole performance of the system, they significantly

enhance their focus area. Among them, research on the reconfiguration for network-on-chip

can be considered significant [53], [59], [81]. A key observation made there is that an NoC

system with reconfigurable topology has greater potential for latency reduction, low power

and area efficiency.

Communication Protocol and Routing Algorithm

Due to the limitation of the area and the device complexity in NoC systems, the com-

munication protocol should be as simple and and as efficiency as possible. Most of the

protocols are somewhat like the open source interconnect (OSI) layer communication pro-

tocol. In other words, they are designed in a layered fashion [19], [58]. Some typical layers

are called application, network control, network and physical. There is no denying that

they are designed to improve the NoC system’s performance, especially for the throughput

and latency. In reference [103], an easy to implement and reusable communication protocol

whose purpose is to reduce the congestion is proposed. It also introduces a kind of adaptive

and deadlock-free routing algorithm. What is more, some of them are also quality of service

(QoS) aware [13], [99]. In [99], the medium access control (MAC) protocol is also collision-

free, and the protocol in [13] is implemented with time division multiplexing (TDM). Other

protocols, such as the one in [2], are designed according to the task structure.

Similar as the communication protocol, routing scheme in NoC systems should also be

simple. The routing schemes can be divided into two types, static and dynamic. Based on

specific conditions, routing algorithms are designed. The main purpose of a routing scheme

42

is to find the possible shortest path between two nodes. Therefore, some papers [86] provide

complex algorithms. In [92], a kind of power-aware adaptive routing scheme is invented. It

is a dynamic XY routing algorithm, which can decrease the power consumption. Since the

load imbalance is also a source that can impact the performance of an NoC system, a kind of

routing algorithm that works on load balancing is proposed in [95], which is also an adaptive

routing scheme. In [56], [66], [67], [68], [91], routing schemes that can increase the fault

tolerance of the whole system are mentioned. While energy efficiency has been considered,

[89] is a study of deadlock-free routing algorithms. For multicasting, there is a specific

routing scheme, which is described in [51]. It is a look-ahead router and the NoC system can

save area by over 50%. Last but not least, a type of region-based routing mechanism [57]

is also designed for NoC. Despite so many routing schemes, for NoC systems the wormhole

router (always including virtual channels) is the most widely used router due to its high

throughput. Therefore, in research that focuses on performance, the wormhole router is the

default.

Wireless Insertion

In a WNoC system, the number of wireless links is not comparable with the number

of nodes. As a result, if we want to use several wireless links simultaneously methods to

allocate wireless links is an important topic. In [9], [99], [62], simulated annealing (SA) is

used to allocate wireless links. The solution sought in SA is the shortest distance in number

of hops between all hub pairs. Another procedure, named evaluation algorithms (EAs) [22],

[62] inserts a limited number of wireless links to a large number of nodes. Another design

[101] uses a combination of simulated annealing and evaluation algorithm for a larger-scale

transit route network optimization. The main difference between SA and EAs is that EAs

can find better solutions while SA can get comparably good result faster [37].

43

8.1.3 Performance

Many researchers give the performance of WNoC systems. However, for systems using

differing topologies, communication protocols, number of functional nodes and other at-

tributes, the performances can be quite different. Here, we point out the best performances

that have been achieved, classified according to the communication method used.

For a WNoC system that has 512 nodes, when CNT-based antennas are used, i.e., there

can be up to 24 simultaneous wireless links, the maximum throughput can be about 0.7 flits

per core per cycle, where flit is the transmission unit that contains 32 bits. For a 256-node

WNoC system, the throughput can also get the same level when the load injection is larger

than 0.7. For the NoC system with the same size, the throughputs are 0.1 flits (512 nodes)

and 0.2 flits (256 nodes) per core per cycle. The best condition of packet energy consumption

in a 512-nodes WNoC system is less than 40nJ when the there are 24 wireless links. For the

same condition, the packet energy consumption in a 256-nodes WNoC system is about 25nJ.

Performance with mm-wave antennas is somewhat different from the CNT-based an-

tenna WNoC system. In a 512-core system, in the best condition, say, mesh topology for

subnet and star-ring topology for top level, the least packet energy is in range 100nJ to

105nJ, which is consumed when the number of wireless interfaces is in the range 2 to 10.

With the same condition (also the best condition), in a 256-core WNoC system, the least

packet energy is about 90nJ when the number of interfaces is in range 2 to 6.

Since the number of wireless links is limited, there must be an upper bound on the total

number of the nodes (referring to functional nodes) in a WNoC system. When the number of

node goes beyond this bound, the performance will not improve, or even decrease as latency

rises. In [9] and [25], the authors indicate that the upper bound is around 512 for a WNoC

system according to the techniques available thus far.

44

8.2 Future Work

To sum up, there are four aspects of WNoC systems that need further development.

These are outlined in this section.

First of all, although the CMOS technology and fabrication techniques are very advanced

and there are so many system-on-chip (SoC) devices already exist, the mm-scale RF devices

for WNoC still have a huge developed gap. To be specific, we need to increase the number of

wireless links that can work simultaneously. This may relate to the manufacturing methods.

Other devices, especially router, are still being studied by many organizations. Again, to be

more specific, for a router designing a superior arbiter, better structure and channel allocating

mechanism and an improved switch function can all contribute to enhanced performance for

the whole WNoC system.

Additionally, for the system level feasible techniques will need be developed for opti-

mizing the WNoC architectures. More potential protocols are waiting to be designed in the

future. The protocols include the communication protocol and topology protocol. At the

topology level, there is no consensus about which is the best, although some popular kinds

and their combinations have been used and studied. Maybe there is a best choice for a

mesh network. However, for some specific tasks, we would need other special and unique

topologies. For communication protocol, though to some extent one can borrow concepts and

techniques from computer networks, WNoC has its own needs: small and simple. Therefore,

it requires the schemes to be simple. Most of the existing communication protocols are lay-

ered, so, in the future, layered and packet-based communication protocol can be a direction

to move. Besides, a requirement of QoS is also to guarantee transmission quality.

Finally, fault tolerance and reconfiguration are two other aspects to be developed. These

may relate to the system architecture and protocols.

To be frank, comparing with an NoC system, although a WNoC system has much

improved performance, the techniques for WNoC are not so mature as they are for NoC

systems. Still, NoC is the dominant research area. Notably however, there is an increasing

45

number of studies on WNoC systems and the relevant techniques for them are becoming

available, gradually.

46

Bibliography

[1] C. Ababei, “Efficient Congestion-Oriented Custom Network-on-Chip Topology Synthesis,”in Proc. International Conf. Reconfigurable Computing and FPGAs (ReConFig), 2010, pp.352–357.

[2] N. Bagherzadeh and M. Matsuura, “Performance Impact of Task-to-Task CommunicationProtocol in Network-on-Chip,” in Proc. Fifth International Conf. Information Technology:New Generations, 2008, pp. 1101–1106.

[3] L. Benini and D. Bertozzi, “Network-on-Chip Architectures and Design Methods,” in InProc. Computers and Digital Techniques, Mar. 2005, pp. 261–272.

[4] L. Benini and G. DeMicheli, “Networks on Chips: A New SoC Paradigm,” Computer, vol. 35,no. 1, pp. 70–78, Jan. 2002.

[5] Booksim. https://nocs.stanford.edu/cgi-bin/trac.cgi/wiki/Resources/BookSim.

[6] A. Bouhraoua, O. Diraneyya, and M. E. Elrabaa, “A Simplified Router Architecture for theModified Fat Tree Network-on-Chip Topology,” in Proc. NORCHIP, 2009, pp. 1–4.

[7] M. Buchanan, “Nexus: Small Worlds and the Groundbreaking Theory of Networks,” inNorton, W. W. Company, 2003.

[8] P. J. Burke et al., “Quantitative Theory of Nanowire and Nanotube Antenna Performance,”IEEE, Trans. Nanotechnology, vol. 5, no. 4, pp. 314–334, July 2006.

[9] K. Chang, S. Deb, A. Ganguly, X. Yu, S. P. Sah, P. P. Pande, B. Belzer, and D. Heo, “Per-formance Evaluation and Design Trade-Offs for Wireless Network-on-Chip Architectures,” inAccepted for publication in ACM Journal on Emerging Technologies in Computing Systems,Sept. 2011.

[10] Y. Chang, C. Chiu, S. Lin, and C. Liu, “On the Design and Analysis of Fault Tolerant NoCArchitecture Using Spare Routers,” in Proc. 16th Asia and South Pacific Design AutomationConference (ASP-DAC), 2011, pp. 431–436.

[11] M. F. Chang et al., “CMP Network-on-Chip Overlaid with Multi-Band RF-Interconnect,”in Proc. IEEE International Symp. High-Performance Computer Architecture (HPCA), Feb.2008, pp. 191–202.

[12] W. Chifeng, H. Wen-Hsiang, and N. Bagherzadeh, “A Wireless Network-on-Chip Design forMulticore Platforms,” in Proc. 19th Euromicro International Conf. Parallel, Distributed andNetwork-Based Processing (PDP), Feb. 2011, pp. 409–416.

[13] N. Concer, A. Vesco, R. Scopigno, and L. P. Carloni, “A Dynamic and Distributed TDM Slot-Scheduling Protocol for QoS-Oriented Networks-on-Chip,” in Proc. IEEE 29th InternationalConf. Computer Design (ICCD), 2011, pp. 31–38.

47

[14] J. Cong, Y. Huang, and B. Yuan, “A Tree-Based Topology Synthesis for on-Chip Network,”in Proc. IEEE/ACM International Conf. Computer-Aided Design (ICCAD), 2011, pp. 651–658.

[15] Cygwin. http://www.cygwin.com/.

[16] W. J. Dally, “Virtual-channel Flow Control,” IEEE Trans. On Parallel and Distributed sys-tem, vol. 3, pp. 194–205, Mar. 1992.

[17] F. Darve, A. Sheibanyrad, P. Vivet, and F. Petrot, “Physical Implementation of an Asyn-chronous 3D-NoC Router Using Serial Vertical Links,” in Proc. IEEE Computer SocietyAnnual Symp. VLSI (ISVLSI), 2010, pp. 25–30.

[18] S. Deb, A. Ganguly, K. Chang, P. P. Pande, B. Belzer, and D. Heo, “Enhancing Performanceof Network-on-Chip Architectures with Millimeter-Wave Wireless Interconnects,” in Proc.IEEE international Conference on ASAP, 2010, pp. 73–80.

[19] M. Dehyadgari, M. Nickray, A. Afzali-Kusha, and Z. Navabi, “A New Protocol Stack Modelfor Network on Chip,” in Proc. IEEE Computer Society Annual Symp. Emerging VLSI Tech-nologies and Architectures, 2006.

[20] J. Draper and F. Petrini, “Routing in Bidirectional k-ary n-cube switch the Red RoverAlgorithm,” in Proc. of the International conference on Parallel and Distributed ProcessingTechniques and Applications, 1997, pp. 1184–1193.

[21] T. Ehsan, T. Mahmoud, and K. Sara, “An Optimized Phased-Array Antenna for Intra-ChipCommunications,” in Proc. Loughborough Antennas and Propagation Conference (LAPC),2011, pp. 1–4.

[22] A. E. Eiben and J. E. Smith, “Introduction to Evolutionary Computing,” in Springer-Verlag,(Berlin, Heidelberg), 2003.

[23] F. Fu, S. Sun, J. Song, J. Wang, and M. YU, “A NoC Performance Evaluation PlatformSupporting Designs at Multiple Levels of Abstraction,” in Proc. 4th IEEE Conference onIndustrial Electronics and Applications, May 2009, pp. 425–429.

[24] M. Fukuda, P. K. Saha, N. Sasaki, and T. Kikkawa, “A 0.18µ CMOS Impulse Radio BasedUWB Transmitter for Global Wireless Interconnections of 3D Staked-Chip System,” in Pro.International Conf. Solid State Devices and Materials, Sept. 2006, pp. 72–73.

[25] A. Ganguly, K. Chang, S. Deb, P. P. Pande, B. Belzer, and C. Teuscher, “Scalable HybridWireless Network-on-Chip Architectures for Multicore Systems,” Computer, vol. 60, no. 10,pp. 1485–1502, Oct. 2011.

[26] C. Grecu, P. P. Pande, A. Ivanov, and R. Saleh, “Timing Analysis of Network on ChipArchitectures for MP-SoC Platforms,” Microelectronics Journal, vol. 36, no. 9, pp. 833–845,Sept. 2005.

[27] W. M. J. Green et al., “Ultra-Compact, Low RF Power, 10Gb/s Silicon Mach-Zehnder Mod-ulator,” Optics Express, vol. 15, no. 25, pp. 17106–17113, 2007.

[28] M. Horowitz and W. J. Dally, “How Scaling Will Change Processor Architecture,” in Proc.International Solid State Circuits Conf. (ISSCC), Feb. 2004, pp. 132–133.

[29] M. A. Horowitz et al., “The Feature of Wires,” Proc. IEEE, vol. 89, no. 4, pp. 490–504, Apr.2001.

48

[30] J. Huang, J. Wu, Y. Chiou, and J. C. F. Jou, “A 24/60GHz Dual-Band Millimeter-Waveon-Chip Monopole Antenna Fabricated with a 0.13µ CMOS Technology,” in Proc. IEEEInternational Workshop Antenna Technology, 2009, pp. 1–4.

[31] Y. Huang, W. Y. Yin, and Q. H. Liu, “Performance Prediction of Carbon Nanotube BundleDipole Antennas,” IEEE Trans. Nanotech., vol. 7, no. 3, pp. 331–337, Sept. 2008.

[32] Y. Huang et al., “Performance Prediction of Carbon Nanotube Bundle Dipole Antennas,”IEEE Trans. Nanotechnology, vol. 7, no. 3, pp. 331–337, May 2008.

[33] D. Hung et al., “Terahertz CMOS Frequency Generator Using Linear Superposition Tech-nique,” IEEE Journal of Solid State Circuits, Dec. 2008.

[34] ITRS. International Technology Roadmap for Semiconductors, 2005 edition.

[35] ITRS, 2007. http://www.itrs.net/Links/2007ITRS/Home2007.htm.

[36] M. Janidarmian, V. S. Bokharaie, A. Khademzadeh, and M. Tavanpour, “Sorena: Newon Chip Network Topology Featuring Efficient Mapping and Simple Deadlock Free Rout-ing Algorithm,” in Proc. IEEE 10th International Conference Computer and InformationTechnology (CIT), 2010, pp. 2290–2299.

[37] T. Jansen and I. Wegener, “A Comparison of Simulated Annealing with a Wimple Evolution-ary Algorithm on Pseudoboolean Functions of Unitation,” in Theoretical Computer Science386, 2007, pp. 73–93.

[38] K. Jensen et al., “Nanotube Radio,” Nano Letter, vol. 7, pp. 3508–3511, Oct. 2007.

[39] K. Jensen et al., “Nanomechanical Radio Transmitter,” Physica Status Solidi B, vol. 245,no. 10, pp. 2323–2325, Sept. 2008.

[40] K. Kempa et al., “Carbon Nanotubes as Optical Antenna,” in Advanced Materials, 2007, pp.421–426.

[41] G. N. Khan and V. Dumitriu, “Throughput-Based Network-on-Chip Topology Generationand Analysis,” in Proc. Canadian Conference Electrical and Computer Engineering, CCECE,2009, pp. 180–184.

[42] J. Kim, J. Balfour, and W. J. Dally, “Flattened Butterfly Topology for On-Chip Networks,”in Proc. 40th Annual IEEE/ACM International Symp. Microarchitecture, MICRO, 2007, pp.172–182.

[43] J. Kim, W. J. Dally, and D. Abts, “Flattened Butterfly: A Cost-Efficient Topology for High-Radix Networks,” in Proc. International Symposium on Computer Architecture (ISCA), 2007,pp. 126–137.

[44] J. Kim, D. Park, C. Nicopoulos, N. Vijaykrishnan, and C. R. Das, “Design and analysis ofan NoC architecture from performance, reliability and energy perspective,” in Architecturefor networking and communications systems, 2005. ANCS 2005. Symposium on, Oct. 2005,pp. 173–182.

[45] S. Kirkpatrick et al., “Optimization by Simulated Annealing,” Scince, vol. 220, pp. 671–680,1983.

[46] T. Krishna et al., “NoC with Near-Ideal Express Virtual Channels Using Global Line Com-munication,” in Proc. IEEE Symp. High Performance Interconnects (HOTI), Aug. 2008, pp.11–20.

49

[47] A. Kumar, L.-S. Peh, and N. K. Jha, “Token Flow Control,” in Proc. 41st IEEE/ACMInternational Symposium on Micro Architecture, 2008, pp. 342–353.

[48] A. Kumar et al., “Toward Ideal On-Chip Communication Using Express Virtual Channels,”IEEE Micro., vol. 28, no. 1, pp. 80–90, Feb. 2008.

[49] B. G. Lee et al., “Ultrahigh-Bandwidth Silicon Photonic Nanowire Waveguides for On-ChipNetworks,” IEEE Photonics Technology Letters, vol. 20, no. 6, pp. 398–400, Mar. 2008.

[50] J. Lee et al., “A low-power fully integrated 60GHz transceiver system with OOK modulationand on-board Antenna assembly,” in Proceedings of IEEE Solid-State Circuits Conference,ISSCC 2009, 2009, pp. 316–317,317a.

[51] W. Lei, P. Kumar, R. Boyapati, H. Y. Ki, and J. K. Eun, “Efficient lookahead routingand header compression for multicasting in networks-on-chip,” in Proc. ACM/IEEE Symp.Architectures for Networking and Communications Systems (ANCS), 2010, pp. 1–10.

[52] J. Lin et al., “Communication Using Antennas Fabricated in Silicon Integrated Circuits,”IEEE Journal of solid-state circuits, vol. 42, no. 8, pp. 1678–1687, Aug. 2007.

[53] A. Logvinenko and D. Tutsch, “A Reconfiguration Technique for Area-Efficient Network-on-Chip Topologies,” in Proc. International Symp. Performance Evaluation of Computer &Telecommunication Systems (SPECTS), 2011, pp. 259–264.

[54] Y. Lu, J. McCanny, and S. Sezer, “Generic Low-Latency NoC Router Architecture for FPGAComputing Systems,” in Proc. International Conf. Field Programmable Logic and Applica-tions (FPL), 2011, pp. 82–89.

[55] P. Magarhack and P. G. Paulin, “System-on-Chip beyond the Nanometer Wall,” Proc. DesignAutomation Conf. (DAC), pp. 419–424, June 2003.

[56] A. Mehranzadeh, A. Khademzadeh, and A. Mehran, “FADyAD - Fault and CongestionAware Routing Algorithm Based on DyAD Algorithm,” in Proc. 5th International Symp.Telecommunications (IST), 2010, pp. 274–279.

[57] A. Mejia, M. Palesi, J. Flich, S. Kumar, P. Lopez, R. Holsmark, and J. Duato, “Region-Based Routing: A Mechanism to Support Efficient Routing Algorithms in NoCs,” IEEETrans. Very Large Scale Integration (VLSI) Systems, vol. 17, pp. 356–369, 2009.

[58] M. Millberg, E. Nilsson, R. Thid, and A. Jantsch, “The Nostrum Backbone-A Communica-tion Protocol Stack for Networks on Chip,” in Proc. 17th International Conf. VLSI Design,2004, pp. 693–696.

[59] M. Modarressi, A. Tavakkol, and H. Sarbazi-Azad, “Application-Aware Topology Reconfig-uration for On-Chip Networks,” IEEE Trans. Very Large Scale Integration (VLSI) Systems,vol. 19, pp. 2010–2022, 2011.

[60] Noxim. http://noxim.sourceforge.net/.

[61] U. Y. Orgras and R. Marculescu, “Its a Small World After All: NoC Performance Opti-mization via Long-Range Link Insertion,” IEEE Trans. Very Large Scale Integration (VLSI)Systems, vol. 14, no. 7, pp. 693–706, July 2006.

[62] P. P. Pande, A. Ganguly, K. Chang, and C. Teuscher, “Hybrid Wireless Network on Chip:A New Paradigm in Multi-Core Design,” in Proc. 2nd International Workshop Network onChip Architectures, NoCArc, Dec. 2009, pp. 71–76.

50

[63] P. P. Pande, C. Grecu, A. Ivanov, and R. Saleh, “Design of a Switch for Network on ChipApplications,” in Proc. International Symp. Circuits and Systems (ISCAS), volume 5, May2003, pp. 217–220.

[64] P. P. Pande, C. Grecu, M. Jones, A. Ivanov, and R. Saleh, “Effect of Traffic Localization onEnergy Dissipation in NoC-based Interconnect,” in Proc. IEEE International Symposium onCircuits and Systems, May 2005, pp. 1774–1777.

[65] P. P. Pande, C. Grecu, M. Jones, A. Ivanov, and R. Saleh, “Performance Evaluation andDesign Trade-Offs for Network-on-Chip Interconnect Architectures,” Computer, vol. 54, no. 8,pp. 1025–1040, Aug. 2005.

[66] S. Pasricha and Z. Yong, “A Low Overhead Fault Tolerant Routing Scheme for 3D Networks-on-Chip,” in Proc. 12th International Symp. Quality Electronic Design (ISQED), 2011, pp.1–8.

[67] S. Pasricha and Z. Yong, “NS-FTR: A Fault Tolerant Routing Scheme for Networks on ChipWith Permanent and Runtime Intermittent Faults,” in Proc. 16th Asia and South PacificDesign Automation Conference (ASP-DAC), 2011, pp. 443–448.

[68] S. Pasricha, Z. Yong, D. Connors, and H. J. Siegel, “OE+IOE: A Novel Turn Model BasedFault Tolerant Routing Scheme for Networks-on-Chip,” in Proc. IEEE/ACM/IFIP Interna-tional Conf. Hardware/Software Codesign and System Synthesis (CODES+ISSS), 2010, pp.85–93.

[69] V. F. Pavlidis and E. G. Friedman, “3D Topologies for Networks-on-Chip,” IEEE Trans.Very Large Scale Integration (VLSI) Systems, vol. 15, no. 10, pp. 1081–1090, Oct. 2007.

[70] K. Payandehjoo and R. Abhari, “Characterization of on-Chip Antennas for Millimeter-WaveApplications,” in Proc. IEEE International Symp. Antennas and Propagation, 2009, pp. 1–4.

[71] R. Qiu, H. Liu, and X. Shen, “Ultra-Wideband for Multiple Acess Communication,” IEEEComm. Magazine, vol. 44, no. 2, pp. 80–87, Feb. 2005.

[72] A.-M. Rahmani, M. Daneshtalab, P. Liljeberg, and H. Tenhunen, “Power-Aware NoC RouterUsing Central Forecasting-Based Dynamic Virtual Channel Allocation,” in Proc. IEEE In-ternational Symposium Circuits and Systems (ISCAS), 2010, pp. 3224–3227.

[73] R. S. Ramanujam, V. Soteriou, B. Lin, and L.-S. Peh, “Design of a High-ThroughputDistributed Shared-Buffer NoC Router,” in Proc. Fourth ACM/IEEE International Symp.Networks-on-Chip (NOCS), 2010, pp. 69–78.

[74] N. Sasaki, M. Fukuda, M. Nitta, K. Kimoto, and T. Kikkawa, “A Single-Chip Ultra-Wideband Receiver Using Silicon Integrated Antennas for inter-Chip Wireless Interconnec-tion,” in Pro. International Conf. Solid State Devices and Materials, Sept. 2006, pp. 70–71.

[75] E. Seok et al., “A 410GHz CMOS Push-Push Oscillator With an on-Chip Patch Antenna,”in Proc. International Solid State Circuits Conf. (ISSCC), 2008.

[76] J. Sepulveda, M. Strum, W. Chau, and G. Gogniat, “A Multi-Objective Approach for Multi-Application NoC Mapping,” in 2011 IEEE Second Latin American Symposium on, Circuitsand Systems (LASCAS), Feb. 2011, pp. 1–4.

[77] A. Shacham et al., “Photonic Network-on-Chip for Future Generations of Chip Multi-Processors,” IEEE Trans. Computers, vol. 57, no. 9, pp. 1246–1260, 2008.

51

[78] M. Shin and J. Kim, “Leveraging Torus Topology with Deadlock Recovery for Cost-Efficienton-Chip Network,” in Proc. IEEE 29th International Conf. Computer Design (ICCD), 2011,pp. 25–30.

[79] D. A. Siguenza-Tortosa and J. Nurmi, “Topology Design for Global Link Optimization inApplication Specific Network-on-Chips,” in Proc. International Symp. System-on-Chip, 2004,pp. 135–138.

[80] E. Sousa and J. Silvester, “Spreading Code Protocols for Distributed Spread-Spectrum PacketRadio Networks,” IEEE Trans. Communications, vol. 36, no. 3, pp. 272–281, Mar. 1988.

[81] M. B. Stensgaard and J. Sparso, “ReNoC: A Network-on-Chip Architecture with Recon-figurable Topology,” in Proc. Second ACM/IEEE International Symp. Networks-on-Chip,NoCS, 2008, pp. 55–64.

[82] C. Teuscher, “Nature-Inspired Interconnects for Self-Assembled Large-Scale Network-on-Chip Designs,” Chaos, vol. 17, no. 2, 2007.

[83] V. M. Vidya, R. Thilagavathy, and M. Bhaskar, “Low Power, High Performance CurrentMode Transceiver for Network-on-Chip Communication,” in Proc. Internattional Conf. Sig-nal Processing, Communication, Computing and Networking Technologies (ICSCCN), 2011,pp. 223–227.

[84] H. Wang, L.-S. Peh, and S. Malik, “A Technology-Aware and Energy-Oriented Topology Ex-ploration for on-Chip Networks,” in Proc. Design, Automation and Test in Europe, volume 2,2005, pp. 1238–1243.

[85] Y. Wang and D. Zhao, “ The Design and Synthesis of a Synchronous and Distributed MACProtocol for Wireless Network-on-Chip,” in Proc. IEEE/ACM International Conf. Computer-Aided Design, 2007, pp. 612–617.

[86] Y. Wang and D. Zhao, “Design and Implementation of Routing Scheme for Wireless Network-on-Chip,” in Proc. IEEE International Symp. Circuits and Systems, ISCAS, 2007, pp. 1357–1360.

[87] C. Wu, S. Chai, Y. Li, and Z. Yang, “Design of a Dual-Switching Mode NOC Router Mi-croarchitecture,” in Proc. International Conf. Electrical and Control Engineering (ICECE),2010, pp. 2733–2736.

[88] C. Wu, Y. Li, S. Chai, and Z. Yang, “Lottery Router: A Customized Arbitral Priority NOCRouter,” Proc. International Conf. Computer Science and Software Engineering, vol. 3, pp.411–418, 2008.

[89] R. Wu, Y. Wang, and D. Zhao, “A Low-Cost Deadlock-Free Design of Minimal-TableRerouted XY-Routing for Irregular Wireless NoCs,” in Proc. Fourth ACM/IEEE Interna-tional Symp, Networks-on-Chip (NOCS), 2010, pp. 199–206.

[90] Y. Xinmin, S. P. Sah, S. Deb, P. P. Pande, B. Belzer, and H. Deukhyoun, “A WidebandBody-Enabled Millimeter-Wave Transceiver for Wireless Network-on-Chip,” in Proc. IEEE54th International Midwest Symp. Circuits and Systems (MWSCAS), 2011, pp. 1–4.

[91] M. Yang, T. Li, Y. Jiang, and Y. Yang, “Fault-Tolerant Routing Schemes in RDT(2,2,1)/-Based Interconnection Network for Networks-on-Chip Design,” in Proc. 8th InternationalSymp. Parallel Architectures,Algorithms and Networks, ISPAN, 2005.

52

[92] S.-G. Yang, L. Li, Y.-A. Zhang, B. Zhang, and Y. Xu, “A Power-Aware Adaptive RoutingScheme for Network on a Chip,” in Proc. 7th International Conf. ASIC, ASICON, 2007, pp.1301–1304.

[93] B. Zafar, J. Draper, and T. M. Pinkston, “Cubic Ring Networks: A Polymorphic Topologyfor Network-on-Chip,” in Proc. 39th International Conf. Parallel Processing (ICPP), 2010,pp. 443–452.

[94] J. Zhang, H. Gu, and Y. Yang, “A High Performance Optical Network on Chip Based on ClosTopology,” Proc. 2nd International Conf. Future Computer and Communication (ICFCC),vol. 2, pp. 63–68, 2010.

[95] Y. Zhang and Q. Cao, “RIPNoC: A Distributed Routing Scheme for Balancing on-ChipNetwork Load,” in Proc. Asia Pacific Conf. Postgraduate Research in Microelectronics andElectronics (PrimeAsia), 2010, pp. 351–355.

[96] Y. Zhang and D. Zhao, “Design and Implementation of Routing Scheme for Wireless Network-on-Chip,” in Proc. IEEE International Symp. Circuits and Systems, May 2007, pp. 1357–1360.

[97] D. Zhao and Y. Wang, “MTNET: Design and Optimization of a Wireless SoC Test Frame-work,” in Proc. IEEE international SoC Conf., Sept. 2006, pp. 239–242.

[98] D. Zhao and Y. Wang, “Feasibility Investigation of RF/Wireless Technology for Intra-/Inter-Chip Communication,” in Technical Report TR-007-8-002, Center for Advanced ComputerStudies, Univ. of Louisiana at Lafayette, 2007.

[99] D. Zhao and Y. Wang, “SD-MAC: Design and Synthesis of a Hardware-Efficient Collision-Free QoS-Aware Protocol for Wireless Network-on-Chip,” 2008.

[100] D. Zhao, Y. Wang, J. Li, and T. Kikkawa, “Design of Multi-Channel Wireless NoC to ImproveOn-Chip Communication Capacity,” in Proc. Fifth IEEE/ACM International Symposium onNetworks on Chip (NoCS), May 2011, pp. 177–184.

[101] F. Zhao and J. X. Zeng, “Simulated Annealing-Genetic Algorithm for Transit Network Opti-mization,” in Compo. in Civ. Engrg. 20, 57 (2006), DOI:10.1061/(ASCE)08873801(2006)20:1(57).

[102] H. Zhou, X. Chen, D. S. Espinoza, A. Mickelson, and D. S. Filipovic, “Nanoscale OpticalDielectric Rod Antenna for On-Chip Interconnecting Networks,” IEEE Trans. MicrowaveTheory and Techniques, vol. 59, pp. 2624–2632, 2011.

[103] L. Zi-di and J. Lin, “A Congestion Avoidance Communication Protocol for Network onChip,” in Proc. IEEE International Conf. Intelligent Computing and Intelligent Systems(ICIS), volume 2, 2010, pp. 88–92.

53