ECODANE - Reducing Energy COnsumption in DAta Center NEtworks based on Traffic Engineering

Truong Thu Huong#1, Daniel Schlosser*2, Pham Ngoc Nam#3, Michael Jarschel*4, Nguyen Huu Thanh#5, Rastin Pries*6

# Hanoi University of Science and Technology, Vietnam{1huong.truong,3pnnam-fet, 5thanhnh}@mail.hut.edu.vn

* University of Wuerzburg, Germany{2schlosser, 3michael.jarschel, 6pries}@informatik.uni-wuerzburg.de

POD 0 POD 1 POD 2 POD 3

Core layer

Aggregationlayer

Edge layer

OptimizerCalculate the optimal topology given the

current traffic conditions

RoutingConcentrate traffic on a

minimum number of links

Power ControlAdjust link/port/switch state

Data Center Environment(Switches, servers and links)

Traffic Generator

m

Monitoring

Traffic StateBit rate, packet rate per port

TopologyLink/Port/switch

state

StatisticOpenFlowRequestMessages

Statistic OpenFlowResponse Messages

Mininet

D-ITG, Iperf(log-normal,…)

NetFPGA(OpenFlow)

NetFPGACards

A

NOX

Currently: Hierarchical Load-Balancing Routing

Algorithm

Frequency(Mhz)

1,501,63

1,721,83

2,012,17 2,22

2,54

1,00

1,50

2,00

2,50

3,00

50 60 70 80 90 100 110 125

Power (W)

�� Tools:�� Emulation: Mininet�� Testbed: 4-ary elastic tree based on NetFPGA

OpenFlow switches �� Network components: �� Optimizer: NOX controller gathering network traffic

statistics based on topology-aware heuristics to find minimum power network subset.

�� Power Control: COntrol of power states of network devices (switches, line cards etc.) through OpenFlow messages and Mininet APIs. Power management module integrated on NetFPGA platform.

�� Forwarding: A NOX module to optimize routes based on the reduced topology.

�� Traffic Generator: Able to generate different traffic patterns, gathered from data center traffic measurements. Based on D-ITG

Deployment

Initial Results

��Experiment scenarios:� Near traffic (within rack), mid-traffic (within POD), far-

traffic (global)� Traffic pattern: lognormal

��Results:� Energy saving between 10% - 35% depending on traffic� On NetFPGA, by reducing working clock frequency, energy

consumption reduces significantly

2,53484 2,53541

2,53858

2,54757

2,525

2,53

2,535

2,54

2,545

2,55

No traffic 10Mbps 100Mbps 884Mbps

Traffic load dependent powerPower (W)

��Analyze the impact of traffic volume on the resource and energy consumption in data centers

��Optimize the resource and energy consumption based on data center traffic measurements

��Build an energy-aware network testbed as the basis for green-network research

Objectives of the Work

�� Problems:� High redundancy vs. energy consumption in data center

network topologies� Energy consumption is not inter-related with traffic volume

in current data center networks�� Assumption: � Make use of Elastic-Tree approach based on Fat-Tree

topology

Problems and Assumptions

�� Focus: optimizing power consumption by designing intelligent mechanisms to adapt the set of network components to the total traffic volume

Network Architecture