Greening the Switch

Ganesh Ananthanarayanan and Randy H. Katz University of California, Berkeley Presented By Rajesh Gadipuuri 1

Motivation

• • • Power consumption of Internet equipment is enormous (~$24 billion per year) – Includes switches, end-hosts, servers Efforts to design energy-efficient network equipment – Energy Efficient Ethernet (EEE), Energy Star Reduce power consumption of

network switches

2

Problems and Goals

• • • • Network traffic is observed to be… – Bursty with interspersed idle periods – Diurnal variations

Heavily underutilized network equipment

Theme: Performance vs. power savings The proposed schemes are stand-alone and hence incrementally deployable 3

Outline

•

Switch Architecture

- Power Model - Port Design - Wake-on-Packet - Buffering • • • • - Shadow Ports

Time Window Prediction Power Save Mode Lightweight Alternative Conclusions

4

Switch Architecture – Power Model

• Switch power consumption - Chassis (Power

fixed

)

- Switching fabric (Power

fabric

) - Line card (Power

line-card

) - Ports (Power

port

) • •

Power

switch

Power

port = Power fixed + Power fabric + numLine * Power line-card + numPort *

Schemes concentrate on putting only ports to sleep Total power consumption of ports is 39.4% four line-cards each containing 192 ports) (Cisco Power Calculator, with 5

Port Design [1]

• • Two state power model – High and low power – Transition takes finite time and power Wake-on-Packet – Avoids the overhead of timer-driven transitions to high-powered state during sustained idle periods – Automatically wake up when a packet arrives 6

Port Design [2] – Buffering

• • • Packets need to be buffered when a port is powered down Processes the buffered packets when a port transitions back to high powered state Inbound packets are lost if the port’s circuitry is down (Current design) 7

Port Design [3] – Shadow Ports

• • • • Receives ingress packets if atleast two of the mapped normal ports are powered down Similar hardware as normal ports At least two normal ports need to be powered down simultaneously for power savings Receives only one packet at a time – Simultaneous arrival  Packet Loss 8

Power Reduction Schemes

• • •

Time Window Prediction

- Adaptive Sleep Window - Wake-on-Packet

Power Save Mode

- Adaptive Sleep Window - Wake-on-Packet

Lightweight Alternative

9

Time Window Prediction [1]

-Egress packets that arrive at a port when it is asleep are buffered and sent after the port wakes up -Ingress packets are handled by shadow port and incur no latency Number of packets, N, in the window t

o

Yes N > τ No

Latency Increase

Process packets buffered during t

s

Sleep for time

t s

10

Time Window Prediction [2]

• Adaptive Sleep Window: - TWP is supplied with per-port bound on the tolerable increase in per packet latency - Adapt the sleep time-window (t

s

) to meet the latency bound - Lower bound for sleeping it set to twice the transition time 11

Time Window Prediction [3]

• Wake-on-Packet: - Ports periodically wake up at the end of sleep window - During sustained idle periods, the energy expended due to periodically waking up and staying awake for units t

o

down is significant wasted before powering - If there are no packets in multiple t

o

windows, sleep continuously until a packet arrives 12

Evaluation – Traces

• • • Traces collected from an enterprise network Power reduction schemes produce power savings upto 20 to 35% With the appropriate hardware support in the form WoP, Shadow ports and fast transitioning of the ports between the high and low power states, these power savings reach 90% of optimal algorithm 13

Evaluation [1] – Time Window Prediction

1. Cluster Size vs. Power Savings 2. Cluster Size vs. Packet Loss 14

Evaluation [2] – Time Window Prediction • Power Savings: Shorter t

o

produces higher savings 15

Evaluation [3] – Time Window Prediction • Packet Loss: For buffer sizes greater than 500 KB, packet loss is under 0.25% with WoP 16

Power Reduction Schemes

• • •

Time Window Prediction

- Adaptive Sleep Window - Wake-on-Packet

Power Save Mode

- Adaptive Sleep Window - Wake-on-Packet

Lightweight Alternative

17

Power Save Mode[1]

• • • • • Similar to wireless networks Power Save Mode is primarily based on the switch’s capability to buffer packets The sleep in PSM happens with regularity and is not dependant on the traffic flow Aggressive and periodic sleep, but adaptive Implements Adaptive sleep Window and WoP similar to TWP 18

Evaluation [1] – Power Save Mode

1. Cluster Size vs. Power Savings 2. Cluster Size vs. Packet Loss 19

Evaluation [2] – Power Save Mode

• • Power savings vs. Sleep time window Power savings vs. Latency Bound 20

Evaluation

• Power savings in PSM 21

Power Reduction Schemes

• • •

Time Window Prediction

- Adaptive Sleep Window - Wake-on-Packet

Power Save Mode

- Adaptive Sleep Window - Wake-on-Packet

Lightweight Alternative

22

Lightweight Alternative

• Diurnal patterns in load w.r.t. time of day • Networks are provisioned for peak-loads – Under-utilized during off periods 23

Lightweight Alternative – Solution

• • • • • Time Window Prediction and Power Save mode algorithms – ports Macroscopic view of the traffic as well as switch Lightweight alternative switch for every high-powered switch – Identify slots of low activity – Only one of the two is powered up All machines have connectivity through the high powered switch as well as lightweight alternative The system uses the simple k-Means clustering algorithm to identify slots of low activity 24

Lightweight Alternative – Design Alternatives

• • Lightweight Switch: – Each line card can be substituted by a separate lightweight switch – Integrated switches can be used as lightweight alternatives – Routing tables and other configuration information for the lightweight switch can be transferred from the main switch using protocols like GARP, VLAN registration protocol (GVRP) Wireless: – Connectivity through wireless access point.

25

• • • •

Combining TWP and PSM with Lightweight Alternative

One of the high-powered switch or the lightweight alternative is powered up depending on the prediction for the slot Switches employ either TWP or PSM Assume WoP, port-transition time of 10ms, latency bound of 10ms for TWP and PSM Power savings from Lightweight Alternative together with the TWP and PSM is higher than individual savings 29

Costs

• • • Lightweight Alternative scheme proposes adding extra hardware in the network Average power savings per day is 30% which translates to an economic savings of $37,133 in one year Economic benefits obtained by power savings are clearly higher than the price of the extra hardware (Lightweight Alternative) 30

Power Savings

Power Reduction Scheme

TWP – Adaptive (t 0 = 0.5s) TWP – WoP (t 0 = 0.5s) PSM – Adaptive (t 0 = 0.5s) PSM – WoP (t 0 = 0.5s) LWA LWA combination with TWP LWA combination with PSM Optimal Power Reduction

Power Savings

21.6% 27.3% 19.8% 26.5% 30% 36% 34% 33.9% Themes are evaluated using traces from a Fortune 500 company’s enterprise network of PC clients and file and other servers 31

Conclusions

• • • Switch architecture – shadow port, wake-on packet Power reduction schemes with bounded performance degradations Lightweight alternative is a power-cognizant network architecture 32

Thank You!!!

33