IT Optimization Under Renewable Energy Constraint Gustavo - PowerPoint PPT Presentation

IT Optimization Under Renewable Energy Constraint Gustavo Rostirolla gustavo.rostirolla@irit.fr Advisors: Patricia Stolf and Stephane Caux Stephane Caux, Paul Renaud-Goud, Gustavo Rostirolla and Patricia Stolf. IT Optimization for Datacenters Under Renewable Power Constraint. Euro-Par 2018: Parallel Processing (Turin-Italy).

Agenda • Introduction • IT Optimization Module • Methodology • Results • Conclusion • Current and Future Work Source: www.toulouse-tourisme.com

Introduction Data centers are known as one of the big players when talking about energy consumption; • In 2006, were responsible for 61.4 billion kWh in the United States; • In 2010 about 1.3% of world's electricity; • 2012 2017 18% 16% 34% 47% 15% Devices Devices 29% Networks Networks 20% 21% Data Centers Data Centers Manufacturing Manufacturing Cook, Gary, et al. "Clicking Clean: Who is winning the race to build a Green Internet?." Greenpeace Inc., Washington, DC (2017). � 3

Introduction Shehabi et al. 2016 - United States Data Center Energy Usage Report ~73 billion kWh in 2020 In the last years, the use of cloud computing has been the basis of data • centers, either in a public or private fashion. � 4

Introduction This migration to cloud computing increases the concern about power • utilization, especially when considering on site renewable energy sources and its oscillation over time; Gigawatts Gigawatts World Total 250 227 Gigawatts World Total World Total World Total 500 433 Gigawatts +50 +50 Annual additions Annual Additions +63 +63 +63 200 370 177 177 Capacity Capacity 400 +52 +40 +40 318 +36 138 283 150 +38 +38 +38 300 +45 +45 238 +41 100 198 +29 +29 +39 +39 159 159 100 200 70 +38 +38 121 +30 +30 94 +27 +27 74 40 +20 59 59 100 +15 +15 +15 50 +17 +17 +17 +12 23 16 +8 9 5.1 +6.5 6.7 +2.5 +1.4 +1.4 0 0 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 Adib, Rana, et al. "Renewables 2016 global status report." Global Status Report Renewable Energy Policy Network for the 21st Century (REN21) (2016): 272. Tasks submitted by users needs to be executed inside a time interval • (release time and due date); But: When? Where? At which speed/frequency? • � 5 Resource Time Energy

Introduction - DataZERO IT quality Electrical quality Electrical IT proposal power excess Electrical proposal Power IT power excess Common part 0:00 6:00 12:00 18:00 Time of the day s t a r r u n f r a f c i s n IT quality Electrical quality l t t a r i u u c T r Succesive Negotiation Rounds i c r e I t t u c e r e l E Power Power Decision Decision Module Module Weather H 2 Cell Forecast Negotiation Users 0:00 6:00 12:00 18:00 New tasks Time of the day www.datazero.org IT quality Electrical quality Power Matching power plans � 6 0:00 6:00 12:00 18:00 Time of the day

IT Optimization Module ✤ Set of Jobs J: J j =[t j,i ;pe j,i ;mem j,i ] ✤ t j,i =[t relase j,i ;t duedate j,i ;t durationj,i ] r w d ✤ mem j,i = Memory requested by task j,i ✤ pe j,i = Number of processing elements requested by task j,i ✤ Set of Machines M: M i =[npe i ;mem i ;P min i ;P max i ] Computing Node ✤ [f i ] set of frequencies available Processor 1 Processor 2 Core 1 Core 1 ✤ mem i : Memory available in node Core 2 Core 2 ✤ P min i :Power when node is idle Core 3 Core 3 Core 4 Core 4 ✤ 𝛽 i : Coe ffi cient dependent on the machine Memory ✤ npe i : Number of processing elements ✤ P max i : g(P min i ;f i,l ; 𝛽 i ) Power with processing element at 100% Power ✤ Discrete power curves P available (t): power available at instant t 0 1 2 3 4 5 6 7 89...n � 8 Time

IT Optimization Module Power Metric Task ✤ Output: Violations: 2 Energy: 7 Power (W) t4 ✤ Which task will run where, t1 t2 t3 t5 when, at which frequency; 0 1 2 3 4 5 6 7 8 9...n Time (minutes) ✤ Constraints {Power, CPU, Violations: 1 Energy: 6 Memory} Power (W) t4 ✤ Translated as a set of t3 t5 t1 t2 scheduling possibilities in the 0 1 2 3 4 5 6 7 8 9...n form of a power profile; Time (minutes) Violations: 0 Energy: 6 ✤ Associated with metrics Power (W) (energy, due date violations…) t3 t5 t1 t4 t2 0 1 2 3 4 5 6 7 8 9...n Time (minutes) � 9

IT Optimization Module • Meta-heuristic (Genetic • Greedy Heuristic (Best Fit) : Algorithm) : • Fast scheduling decisions; • Allows to produce a large • Easy implementation; number of adapted solutions; • Tasks can be sorted by • Makes it possible to approach arrival time, due date…; an optimum solution; • Tasks are scheduled in a • Slow execution time; local optimal, limited by the power curve received; • Difficulties in setting parameters (crossover, • The combinations of choices mutation rate, population size, locally optimal do not always selection method). lead to an overall optimum. � 10

IT Optimization Module Genetic Algorithm Selection Start Workflow Crossover Generate Initial Population Mutation Calculate Fitness of Individuals Greedy Algorithm to Set Tasks Start Time Satisfy Yes No End Stop Criterion Greedy Algorithm to adjust DVFS Calculate Fitness of Children Elitism for New Generation � 11

IT Optimization Module Genetic Algorithm: Parents O ff springs Node 3 Node 3 Node 0 Node 2 Node 3 Node 3 Node 2 Node 3 Crossover/Mutation T0 T1 T2 T3 T0 T1 T2 T3 Node 0 Node 0 Node 2 Node 3 Node 0 Node 0 Node 0 Node 2 T0 T1 T2 T3 T0 T1 T2 T3 Greedy Time, Processor and Frequency Node 3 Node 3 Node 0 Node 2 Processor 1 Processor 2 Processor 1 Processor 1 Frequency: F0 Frequency: F1 Frequency: F0 Frequency: F3 Start: 1:00pm Start: 1:00pm Start: 3:00pm Start: 1:00pm End: 2:00pm End: 2:00pm End: 5:00pm End: 2:00pm T0 T1 T2 T3 Node 0 Node 0 Node 2 Node 3 Processor 1 Processor 2 Processor 1 Processor 1 Frequency: F0 Frequency: F1 Frequency: F0 Frequency: F3 Start: 1:00pm Start: 1:00pm Start: 1:00pm Start: 2:00pm End: 2:00pm End: 2:00pm End: 2:00pm End: 3:00pm T0 T1 T2 T3 � 12

IT Optimization Module DVFS: (a) Pr Node 0 Node 0 Node 0 Pr Processor 1 Processor 2 Processor 2 Frequency: F1 Frequency: F0 Frequency: F3 Start: 1:00pm Start: 1:00pm Start: 2:30pm End: 5:00pm End: 2:30pm End: 5:00pm Pr T1 T2 T3 Pr on o ff T1 Processor 1 (b) T2 T3 Processor 2 on o ff T1 Processor 1 (c) T2 T3 Processor 2 � 13

IT Optimization Module • Two execution phases for Genetic Algorithm to improve execution time: • First phase provides an initial placement of tasks respecting a simplified power curve; • Second phase uses the power prediction with all variations to improve this initial placement, allowing the scheduling to take profit of power peaks. 8000 7000 6000 5000 Power (W) 4000 3000 2000 1000 0 0 100 200 300 400 500 600 Time Step Refined Simplified � 14

Evaluation

Evaluation ✤ Computing Resources: ✤ Based in Paravance and Taurus (Grid5000) ✤ 30 Nodes x 2 Processors x 5 Frequencies (1.2 to 2.4 Ghz); ✤ + Overhead of turning on/off a node. T. Mudge, “Power: A first-class architectural design constraint,” Computer, vol. 34, pp. 52– 58, 2001. Source: Villebonnet, V. (2016). Scheduling and Dynamic Provisioning for Energy Proportional � 16 Heterogeneous Infrastructures (Doctoral dissertation, Université de Lyon).

Evaluation Profile II Summer Profile 2 days simulation with DCWorms: ✤ 8000 7000 6000 3 power profiles (real traces) ✤ 5000 Power (W) 4000 3 workloads (Google based) ✤ 3000 2000 234, 569 and 1029 tasks 1000 ✤ 0 (known at beginning of 0 20000 40000 60000 80000 100000 120000 140000 160000 180000 Time (seconds) execution) Sun Wind Total Profile I Profile III Mixed Profile Winter Profile 8000 8000 7000 7000 6000 6000 5000 5000 Power (W) Power (W) 4000 4000 3000 3000 2000 2000 1000 1000 0 0 0 20000 40000 60000 80000 100000 120000 140000 160000 180000 0 20000 40000 60000 80000 100000 120000 140000 160000 180000 Time (seconds) Time (seconds) � 17 Sun Wind Total Sun Wind Total

Results � 18

Results Best Fit Due Date Best Fit Arrival Best Fit Size Genetic Algorithm Genetic Algorithm MO 192 171 200 136 Due Date Violations 124 150 119 100 Violations 58 50 22 19 18 12 12 11 11 6 5 6 4 4 4 3 3 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 234 Tasks 569 Tasks 1029 Tasks 234 Tasks 569 Tasks 1029 Tasks 234 Tasks 569 Tasks 1029 Tasks Profile I Profile II Profile III 78.11 81.03 73.18 77.24 81.24 73.48 77.03 81.39 73.15 90.00 68.14 67.87 67.65 67.30 67.49 67.03 Energy Consumption (kWh) 53.92 53.78 53.26 50.19 50.32 50.38 47.04 46.53 46.12 45.35 45.41 45.63 45.07 45.21 45.15 60.00 Energy 18.34 18.09 17.47 15.81 15.86 15.69 15.40 14.84 14.62 14.79 14.91 14.69 14.94 14.85 14.55 30.00 0.00 234 Tasks 569 Tasks 1029 Tasks 234 Tasks 569 Tasks 1029 Tasks 234 Tasks 569 Tasks 1029 Tasks Profile I Profile II Profile III � 19