study of quality of service >> issues in...

>>

>>Luminous Networks/ Sanjay K. Agrawal, March 5, 2001. IEEE 802.17

Study of Quality of Service Issues in RPR

Sanjay K. Agrawal

Luminous [email protected]

>> Luminous Networks/ Sanjay K. Agrawal, March 5, 2001. IEEE 802.172

Objectives

l Type of service requirement in provider networks in metro and core

l Study of congestion control and fairness requirements and solutions

l RPR framework proposal


Provider Network Type of Service Requirements

l Legacy leased lines

l Voice over IP

l Video services

l Committed Access Rate data service (VLL)

l Over Committed Access Rate data services

l Best effort services


Class of Servicel Time sensitive committed class

u Legacy leased line, Voice over IP, Protected, No degradation on protection

l Time Sensitive committed classu Video, Protected, No degradation on protection

l Time insensitive committed classu Committed data Services, Protected, No degradation on

protection

l Time insensitive over-committed classu Over committed data services, degradation on protection

l Best Effort Serviceu Best effort data services, degradation on protection


Class of Service support

l Bandwidth provisioning maintained on per class basis

l Traffic from one class should not effect the traffic in the other class.


Issue of Congestion & Fairness

l Only necessary in the over-committed class.


The Problem of Congestion

l Traffic sources: TCP linearly increase their bandwidth usage till resources are exhausted. On packet drop they back off exponentially

l Uncontrolled congestion seriously degrades performance

u Buffers fill up

u Packets are dropped, resulting in re-transmissions

u Result: more packet loss and increased latency

u Problem builds until throughput collapses

u 35% link utilization on MAN and WAN links [Caida: www.caida.org]

Th

rou

gh

pu

t

Congestion

Controlled CongestionControlled CongestionControlled Congestion

Uncontrolled CongestionUncontrolled CongestionUncontrolled Congestion


Problem of Fairness

l Per node Fairness in the rings

RPR Ring

Node 0

Node 1

Node 2

Node 3

Node N


Problem of Fairness (cont..)

l Per customer/subscriber fairness in the rings

RPR Ring

Node 0

Node 1

Node 2

Node 3

Node N

Customer 0

Customer 1

Customer 2 Customer 3Customer 4

Customer 9Customer 8

Customer 7

Customer 6

Customer 5



l Per domain fairness in the rings in Public Transparent LAN services

RPR Ring

Node 0

Node 1

Node 2

Node 3

Node N

Domain0

Domain2

Domain1

Domain1

Domain1

Domain2

Domain2

Domain0

Domain0



l Per MPLS aggregate fairness in the rings and mesh of rings.

MPLS Label 0

MPLS Label 1

MPLSLabel 2


Congestion Control & Fairnessl Per Flow Queuing

uATM environment

uDeterministic QoS per flow

uSerious scaling issues in IP networks

l Per Aggregate Flow Queuing

uAggregates based on VPN domain, MPLS labels, nodes, customers.

uDeterministic QoS per aggregate flow

uScaling issue. 1k customers/port, 12 port/line card, 12 line cards/ chassis, 128 chassis in a ring.

uScheduling and congestion control very difficult


Congestion Control & Fairness

l Per class queuing

uScales very nicely

uDeterministic and statistical QoS per aggregate

uScheduling and congestion control relatively simple


Random Early Detection (RED)

uRED:

nAnticipates congestion

nSlows down traffic before queue overflows

nAvoids TCP oscillations

nMaximizes throughput

uRED uses selective packet loss to signal TCP to slow down

unew RED, Blue

RED

Random Early DetectionP

acke

t D

rop

P

rob

abili

ty

Queue Length

“Slope” is adjustable

Queue Max

Pac

ket

Dro

p

Pro

bab

ility

Queue Length Queue Max

Pac

ket

Dro

p

Pro

bab

ility

Queue Length

Standard Service

Queue Max

WithoutRED

WithRED

WithWRED

Premium Service

Std. Min. Prem. Min.

350792_04F7_c1


Problem of Congestion Control & Fairness

l Fairness across flows

l Fairness across responsive and non responsive flows

l Fairness across round trip times (RTT)

l Weighted fairness across aggregates


Buffering vs. BCN

l Backward Congestion Notification (BCN)

u Avoids buffering in the intermediate nodes in the rings.

u Propagates congestion to source nodes.

u Flow control signaling frequency and span distances may be issue

u Interaction with upper layer protocols (TCP or any adaptive) may be issue.

l Buffering

u Avoids internode signaling.

u Well tested

u Requires 50-100ms buffering


Simplify MAC

Optical Transmission Choice(Ethernet, SONET,…new ones)

Ring Operations(Frame Format, Topology Discovery,

Congestion-messaging, Protection)

Service Intelligence(QoS: Policy, Fairness and BW based Forwarding

& Congestion Control)

Vendor Specific

802.17 Specific

PHY Specific

Data TDM Video

Bound Scope

Data PathTopology, Congestion

Error Mgmt Data Path


Proposed RPR MAC Implementation

System QoS

ProtectionSwitching

TopologyDiscovery

RPR MAC

Phy for EastRing

Phy for WestRing

MAC forWEST

MAC forEAST

Traffic Given to RPR


Proposed RPR System Architecture

TopologyAlgorithm

ProtectionAlgorithm

FairnessAlgorithm

TopoTable

Drop Traffic

WestScheduler

EastScheduler

Class BasedQueuing

Class BasedQueuing

WestPHY

EastPHY

Ring Congestion

Add, E-Hop-count, W-Hop-count

Passnon-EF

RPR System

RPRMACLayer

Pass non-EF

D

P

A

A

Transit Buffer

Transit Buffer

Mux

MuxBCN

BCN

Drop

DropNon EF

Pass

Non EFPass

Congestion Control

Add Traffic


Conclusion

l Many service scenarios

l QoS above RPR MAC layer

l Proposal not tied to a particular implementation that addresses only a set of needs.

l Doesn’t preclude services of present and future

l Open to innovation and evolution

l Allows vendor differentiation while insuring interoperability

study of quality of service >> issues in...

Documents