Anshul Gandhi (Carnegie Mellon University) Varun Gupta (CMU), Mor Harchol-Balter (CMU) Michael Kozuch (Intel, Pittsburgh) Requests Server Farm 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) Requests Server Farm How many servers, given request rate ? Don’t want to waste power 1. Server farm model 2. Provisioning for fixed arrival rate 3. Provisioning for unpredictable, time-varying arrival rate 4. Future work BUSY BUSY BUSY IDLE IDLE servers consume a lot of power ~ 60 % of BUSY IDLE OFF OFF 5 BUSY BUSY Turn IDLE servers OFF to save power BUSY OFF HOWEVER OFF OFF OFF 6 To turn on an OFF server .. OFF SETUP BUSY Time delay (setup time) • 1 min – 5 mins and Power penalty • peak power during setup time 7 To turn on an OFF server .. OFF SETUP BUSY Should we ever turn servers OFF ? 8 ON Server states: BUSY PBUSY 240 W IDLE PIDLE 150 W OFF POFF 0W SETUP PSETUP 240 W Setup times: TOFF→ON TON→OFF 200 s 0s Intel Xeon E5320 • 2 X 1.86 GHz quad-core • 4GB memory 9 Requests FCFS Server Farm 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 10 Interested in response time and power conumption Perf/W = 1/(Mean RT X Mean Power) Maximize Perf/W 11 1. Server farm model 2. Provisioning for fixed arrival rate 3. Provisioning for unpredictable, time-varying arrival rate 4. Future work Poisson arrivals Server Farm Max. Perf/W Existing solutions: prediction based, reactive controllers. Is there a simple, yet, near-optimal solution ? 13 Keep n servers always ON (M/M/n) Servers are BUSY or IDLE n* 14 15 Turn servers OFF when IDLE Servers are BUSY, OFF or in SETUP n* Auto-scales if n is high 16 17 TON→OFF < γ E[S]/√ρ 18 Best of {NEVEROFF, INSTANTOFF} is optimal for single-server Multi-server ? For ρ > 10, we are within 20% of OPT 19 1. Server farm model 2. Provisioning for fixed arrival rate 3. Provisioning for unpredictable, time-varying arrival rate 4. Future work Data center demand has daily variations INSTANTOFF can auto-scale 21 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 22 Like INSTANTOFF, except we wait for twait seconds before turning IDLE servers OFF Routing ? MRB routing is crucial ! 23 Rule of thumb: twait ∙ PIDLE = TOFF→ON ∙ PON 24 Worse at higher frequencies 25 1998 World Cup Soccer trace (ITA) 26 1. Server farm model 2. Provisioning for fixed arrival rate 3. Provisioning for unpredictable, time-varying arrival rate 4. 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 28 Anshul Gandhi, Varun Gupta, Mor Harchol-Balter, Michael Kozuch Optimality analysis of energy-performance trade-off for server farm management, PERFORMANCE 2010 Anshul Gandhi, Mor Harchol-Balter, Ivo Adan Server farms with setup costs, PERFORMANCE 2010 Anshul Gandhi, Varun Gupta, Mor Harchol-Balter, Michael Kozuch Energy-efficient dynamic capacity provisioning in server farms, CMU technical report CMU-CS-10-108 29