power-efficient server provisioning in server farms
DESCRIPTION
Anshul Gandhi (Carnegie Mellon University) Varun Gupta (CMU), Mor Harchol-Balter (CMU) Michael Kozuch (Intel, Pittsburgh). Power-efficient server provisioning in server farms. Motivation. Server farms are important for today’s IT infrastructure (Amazon, Google, IBM, HP, …) - PowerPoint PPT PresentationTRANSCRIPT
![Page 1: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/1.jpg)
Power-efficient server provisioning in server farms
Anshul Gandhi (Carnegie Mellon University)
Varun Gupta (CMU), Mor Harchol-Balter (CMU)Michael Kozuch (Intel, Pittsburgh)
![Page 2: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/2.jpg)
Motivation
Server farms are important for today’s IT infrastructure (Amazon, Google, IBM, HP, …)
However, server farms cost a lot of money to power ($4 billion in 2006)
Server FarmRequests
![Page 3: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/3.jpg)
High-level problem statement
How many servers, given request rate ? Don’t want to waste power
RequestsServer Farm
![Page 4: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/4.jpg)
Outline
1. Server farm model
2. Provisioning for fixed arrival rate
3. Provisioning for unpredictable, time-varying arrival rate
4. Future work
![Page 5: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/5.jpg)
5
Server farms
IDLE servers consume a lot of power
~ 60 % of BUSY
BUSY
BUSY
BUSY
IDLE
IDLE
OFF
OFF
![Page 6: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/6.jpg)
6
Server farms
Turn IDLE servers OFF to save power
BUSY
BUSY
BUSY
OFF
OFF
OFF
OFF
HOWEVER
![Page 7: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/7.jpg)
7
Setup cost
To turn on an OFF server ..
BUSYOFF SETUP
Time delay (setup time)• 1 min – 5 mins
and
Power penalty • peak power during setup time
![Page 8: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/8.jpg)
8
Setup cost
To turn on an OFF server ..
BUSYOFF SETUP
Should we ever turn servers OFF ?
![Page 9: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/9.jpg)
9
Server model
Server states:BUSY PBUSY 240 WIDLE PIDLE 150 WOFF POFF 0 WSETUP PSETUP 240 W
Setup times:TOFF→ON 200 sTON→OFF 0 s
Intel Xeon E5320• 2 X 1.86 GHz quad-core• 4GB memory
ON
![Page 10: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/10.jpg)
10
Server farm model
Poisson arrival process: λ(t) requests/sec Exponentially distributed job sizes: E[S] secs Load: ρ(t) = λ(t) E[S]∙
Minimum # servers to handle incoming load
RequestsFCFS Server Farm
![Page 11: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/11.jpg)
11
Metric
Interested in response time and power conumption
Perf/W = 1/(Mean RT X Mean Power)
Maximize Perf/W
![Page 12: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/12.jpg)
Outline
1. Server farm model
2. Provisioning for fixed arrival rate
3. Provisioning for unpredictable, time-varying arrival rate
4. Future work
![Page 13: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/13.jpg)
13
Provisioning for fixed arrival rate
Existing solutions: prediction based, reactive controllers.
Is there a simple, yet, near-optimal solution ?
Poisson arrivals
Server Farm
Max. Perf/W
![Page 14: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/14.jpg)
14
NEVEROFF
Keep n servers always ON (M/M/n) Servers are BUSY or IDLE
*n
![Page 15: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/15.jpg)
15
Perf/W for NEVEROFF
![Page 16: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/16.jpg)
16
INSTANTOFF
Turn servers OFF when IDLE Servers are BUSY, OFF or in SETUP
*n
Auto-scales if n is high
![Page 17: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/17.jpg)
17
Perf/W for INSTANTOFF
![Page 18: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/18.jpg)
18
NEVEROFF vs. INSTANTOFF
TON→OFF < γ E[S]/√ρ
![Page 19: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/19.jpg)
19
Near-optimality
Best of {NEVEROFF, INSTANTOFF} is optimal for single-server
Multi-server ?
For ρ > 10, we are within 20% of OPT
![Page 20: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/20.jpg)
Outline
1. Server farm model
2. Provisioning for fixed arrival rate
3. Provisioning for unpredictable, time-varying arrival rate
4. Future work
![Page 21: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/21.jpg)
21
Unpredictable, time-varying demand
Data center demand has daily variations
INSTANTOFF can auto-scale
![Page 22: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/22.jpg)
22
Unpredictable, time-varying demand
NEVEROFF requires continual updates based on predicted load
Predictions are not always accurate
Can we find a simple traffic-oblivious policy? Auto-scaling in nature
![Page 23: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/23.jpg)
23
DELAYEDOFF
Like INSTANTOFF, except we wait for twait seconds before turning IDLE servers OFF
Routing ?
MRB routing is crucial !
![Page 24: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/24.jpg)
24
twait
Rule of thumb: twait P∙ IDLE = TOFF→ON P∙ ON
![Page 25: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/25.jpg)
25
Near-optimality
Worse at higher frequencies
![Page 26: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/26.jpg)
26
Auto-scaling capabilities
1998 World Cup Soccer trace (ITA)
![Page 27: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/27.jpg)
Outline
1. Server farm model
2. Provisioning for fixed arrival rate
3. Provisioning for unpredictable, time-varying arrival rate
4. Future work
![Page 28: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/28.jpg)
28
Future work
Experimental evaluation of proposed schemes Preliminary experiments on 15-server testbed using
CPU-bound workload and sinusoidal arrival pattern Experimental results agree with analysis Web workloads:▪ What does the experimental setup look like ?
Try out various arrival traces and workloads
![Page 29: Power-efficient server provisioning in server farms](https://reader035.vdocuments.us/reader035/viewer/2022062323/56815e35550346895dcc9847/html5/thumbnails/29.jpg)
29
Thank You!
Anshul Gandhi, Varun Gupta, Mor Harchol-Balter, Michael KozuchOptimality analysis of energy-performance trade-off for server farm management, PERFORMANCE 2010
Anshul Gandhi, Mor Harchol-Balter, Ivo AdanServer farms with setup costs, PERFORMANCE 2010
Anshul Gandhi, Varun Gupta, Mor Harchol-Balter, Michael KozuchEnergy-efficient dynamic capacity provisioning in server farms, CMU technical report CMU-CS-10-108