Energy Aware Network Operations Authors: Priya Mahadevan, Puneet Sharma, Sujata Banerjee, Parthasarathy Ranganathan HP Labs IEEE Global Internet Symposium.

Energy Aware Network Operations

Authors: Priya Mahadevan, Puneet Sharma, Sujata Banerjee, Parthasarathy Ranganathan

HP LabsIEEE Global Internet Symposium (April 2009)

Speaker: Sookhyun Yang

Introduction

2

• This paper analyzes three energy-saving schemes to configure a switch in a data center network with link redundancy.

[Data center network topology]

Energy Saving in a Switch• Linear power model

– Power consumed by each switch depends on • Number of active ports • Line capacity of a port

• How can we control a switch for saving energy?– Disable a switch port– Dynamically adapt a port’s link

capacity based on its load– Turn off a line-card that have no

active ports– Power off a switch that is not

being used

3

[A switch with line-cards]Line-card

port

Three Energy Saving Schemes• Centralized approach

• LSA (Link State Adaptation)– Adapts a port’s link capacity (disabled,

10Mbps, 100Mbps, and 1Gbps) according to link utilization.

• NTC (Network Traffic Consolidation)– Consolidates traffic into fewer links

(and switches).

– Disables unused links (and switches).

• SLC (Server Load Consolidation)– Migrates jobs for minimizing the

number of servers being used.

– Applies NTC schemes.

4

[Data center network topology]

Variations of the Three Schemes

• SL (Service Level) Awareness– Adds a constraint to ensure that a link’s utilization never

exceeds a certain threshold.

– Ensures that at least one redundant link exists.

• SL awareness policy is combined with each of the three schemes.

5

Simulation Set-up: Workload

• Simulation based on workload of observed traffic

• 292 web-servers for 5days in April 2008

• System configuration of 292 web-severs– Quad-core processors, two 1Gbps network cards

– Different RAM sizes: 193 servers have 4 GB RAM, 69 servers have 8GB RAM, and 30 servers have 16GB RAM.

• Observed results– Workload is memory-sensitive

• 130 servers use 90% or greater amount of memory, 64 servers use less than 40%.

– Both network bandwidth and CPU of all servers are utilized at most 10%.

6

7[Data center network topology]

292 web servers

26 rack switches

(48 ports per switch)

Simulation Set-up: Network Topology

2 tier-2 switch(6 line card per switch,24 ports per line card)

Simulation: How to Compute Power Consumption

8

Line-card with no active ports

Line speed

• Power consumption

• Perfect knowledge of an oracle – Link utilization

– Job’s traffic specification

– Network topology

LSA (Link State Adaptation)• Link characteristics from 5-day measurement

– 90% of links (light traffic) can be set 10 or 100Mbps.

– Less than 5% of links (heavy traffic) need to be set 1Gbps.

• LSA’s distribution of link speeds between rack and tier-2 switches

9

LSA vs. SL-aware LSA

10

Simulation Results

11

LSA (Link state adaptation)

NTC (Network

traffic consolidatio

n)

SLC (Server load consolidation)

Deployment Issues

• Track traffic workload– Adapt link capacity based on link utilization statistics

– Predict incoming/outgoing traffic

• Transition time for adapting link-speeds is between 1-3 seconds, which can affect network performance.– Buffer or ensure the existence of back-up paths

12

Q&AQ&A

Thank you!

13