Energy-efficient data centers: Exploiting knowledge about application and resources

96e2d41af0b2b857d8ec7cd65d170e72?s=47 GreenLSI
February 03, 2014

Energy-efficient data centers: Exploiting knowledge about application and resources

Presentation by Jose M. Moya at the IEEE Region 8 SB & GOLD Congress (25 – 29 July, 2012).

The current techniques for data center energy optimization, based on
efficiency metrics like PUE, pPUE, ERE, DCcE, etc., do not take into
account the static and dynamic characteristics of the applications and
resources (computing and cooling). However, the knowledge about the
current state of the data center, the past history, the resource
characteristics, and the characteristics of the jobs to be executed
can be used very effectively to guide decision-making at all levels in
the datacenter in order to minimize energy needs. For example, the
allocation of jobs on the available machines, if done taking into
account the most appropriate architecture for each job from the
energetic point of view, and taking into account the type of jobs that
will come later, can reduce energy needs by 30%.

Moreover, to achieve significant reductions in energy consumption of
state-of-the-art data centers (low PUE) is becoming increasingly
important a comprehensive and multi-level approach, ie, acting on
different abstraction levels (scheduling and resource allocation,
application, operating system, compilers and virtual machines,
architecture, and technology), and at different scopes (chip, server,
rack, room, and multi-room).

96e2d41af0b2b857d8ec7cd65d170e72?s=128

GreenLSI

February 03, 2014
Tweet

Transcript

  1. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Energy-efficient data centers:

    Exploiting knowledge about application and resources José  M.  Moya  <jm.moya@upm.es>   Integrated  Systems  Laboratory   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   1  
  2. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Data centers José

     M.Moya  |    Madrid  (Spain),  July  27,  2012   2  
  3. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE José  M.Moya  |

       Madrid  (Spain),  July  27,  2012   3  
  4. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Power distribution José

     M.Moya  |    Madrid  (Spain),  July  27,  2012   4  
  5. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Power distribution (Tier

    4) José  M.Moya  |    Madrid  (Spain),  July  27,  2012   5  
  6. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Contents •  MoHvaHon

      •  Our  approach   –  Scheduling  and  resource   management   –  Virtual  machine   opHmizaHons   –  Centralized  management   of  low-­‐power  modes   –  Processor  design   •  Conclusions   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   6  
  7. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Motivation José  M.Moya

     |    Madrid  (Spain),  July  27,  2012   7   •  Energy  consumpHon  of  data  centers   –  1.3%  of  worldwide  energy  producHon  in  2010   –  USA:  80  mill  MWh/year  in  2011  =  1,5  x  NYC   –  1  data  center  =  25  000  houses   •  More  than  43  Million  Tons  of  CO2  emissions  per   year  (2%  worldwide)   •  More  water  consumpHon  than  many  industries   (paper,  automoHve,  petrol,  wood,  or  plasHc)                      Jonathan  Koomey.  2011.  Growth  in  Data  center  electricity  use  2005  to  2010  
  8. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Motivation José  M.Moya

     |    Madrid  (Spain),  July  27,  2012   8   •  It  is  expected  for  total  data   center  electricity  use  to   exceed  400  GWh/year  by   2015.   •  The  required  energy  for   cooling  will  conHnue  to  be  at   least  as  important  as  the   energy  required  for  the   computaHon.   •  Energy  opHmizaHon  of  future   data  centers  will  require  a   global  and  mulH-­‐disciplinary   approach.   0   5000   10000   15000   20000   25000   30000   35000   2000   2005   2010   World  server  installed  base   (thousands)   High-­‐end  servers   Mid-­‐range  servers   Volume  servers   0   50   100   150   200   250   300   2000   2005   2010   Electricity  use     (billion  kWh/year)   Infrastructure   CommunicaHons   Storage   High-­‐end  servers   Mid-­‐range  servers   Volume  servers   5,75  Million  new  servers  per  year   10%  unused  servers  (CO2  emissions   similar  to  6,5  million  cars)  
  9. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Temperature-dependent reliability problems

    José  M.Moya  |    Madrid  (Spain),  July  27,  2012   9   Time-­‐dependent   dielectric-­‐ breakdown  (TDDB)   ElectromigraHon  (EM)   Stress   migraHon  (SM)   Thermal   cycling  (TC)   ✔   ✖   ✖   ✖  
  10. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Cooling a data

    center José  M.Moya  |    Madrid  (Spain),  July  27,  2012   10  
  11. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE •  VirtualizaHon  

        -­‐  27%   •  Energy  Star  server   conformance       =  6.500   •  Bejer  capacity   planning     2.500   Server improvements José  M.Moya  |    Madrid  (Spain),  July  27,  2012   11   . . . . . . . . . . . . . . . . . . . . . . . . . . . Servidores 5,75 millones de nuevos servido- res se instalan cada año para mantenerse al ritmo de creci- miento de los servicios on-line, y todavía aproximadamente el 10% de los servidores instalados no se utilizan debido a sobrestima- ciones conservadoras a la hora de planificar las necesidades de almacenamiento. La energía utilizada para los servidores en desuso podría compensar las emisiones de 6,5 millones de coches. de los servidores. En un centro de datos convencional, algunas de estas mejoras pueden reducir su impacto en las siguientes cifras: (excluyendo el software) $ $ $ $ A menudo los servidores se sobredimensionan para afrontar picos de demanda, lo que significa que como media suelen funcionar sólo al 20% 2.500 6,5 millones = 6.500 El equivalente a retirar 6.500 coches de las carreteras, me- diante la utilización de servidores acordes a Energy Star, lo que re- duciría el consumo eléctrico de los centros de datos en 82.000 megavatios-hora. 10% no se utilizan US 27% Reducir un 27% el consumo energético mediante la virtuali- zación, lo que reduce la capaci- dad productiva no empleada. El equivalente a la energía consumida por 2.500 hogares en EEUU, mediante una mejor planifica- ción de la capacidad.
  12. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Cooling improvements • 

    Improvements  in  air  flow  management  and   wider  temperature  ranges   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   12   Energy  savings   up  to  25%   25.000   Return  of  investment   in  only  2  years    
  13. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE AC  è  DC

      – 20%  reducHon  of  power  losses  in  the   conversion  process   – 47  million  dollars  savings  of  real-­‐state  costs   – Up  to  97%  efficiency,  energy  saving  enough  to   power  an  iPad  during   70   million  years   Infrastructure improvements José  M.Moya  |    Madrid  (Spain),  July  27,  2012   13  
  14. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE   

           Best practices José  M.Moya  |    Madrid  (Spain),  July  27,  2012   14  
  15. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE And… what about

    IT people? José  M.Moya  |    Madrid  (Spain),  July  27,  2012   15  
  16. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE PUE Power Usage

    Effectiveness •  State  of  the  Art:    PUE  ≈  1,2   –  The  important  part  is  IT  energy  consumpHon   –  Current  work  in  energy  efficient  data  centers  is   focused  in  decreasing  PUE   –  Decreasing  PIT     does  not  decrease  PUE,  but  it  is  seen  in   the  electricity  bill   •   But  how  can  we  reduce  PIT   ?   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   16   !"# = 1 !"#$ = !!"!#$ !!" =! ! !!!!!!!!!!= !!"#$% + !!""#$%& + !!"# !!"#$% ≈ !!""#$%& + !!" !!" !
  17. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Potential energy savings

    by abstraction level José  M.Moya  |    Madrid  (Spain),  July  27,  2012   17  
  18. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Our approach • 

    Global  strategy  to  allow  the  use  of  mulHple   informaHon  sources  to  coordinate  decisions  in  order   to  reduce  the  total  energy  consumpHon   •  Use  of  knowledge  about  the  energy  demand   characterisHcs  of  the  applicaEons,  and     characterisHcs  of  compuEng  and  cooling  resources   to  implement  proacEve  opEmizaEon  techniques   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   18  
  19. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Holistic approach José

     M.Moya  |    Madrid  (Spain),  July  27,  2012   19   Chip   Server   Rack   Room   MulE-­‐ room   Sched  &  alloc   2   1   app   OS/middleware   Compiler/VM   3   3   architecture   4   4   technology   5  
  20. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE 1. Room-level resource

    management José  M.Moya  |    Madrid  (Spain),  July  27,  2012   20   Chip   Server   Rack   Room   MulE-­‐ room   Sched  &  alloc   2   1 app   OS/middleware   Compiler/VM   3   3   architecture   4   4   technology   5  
  21. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Leveraging heterogeneity • 

    Use  heterogeneity  to  minimize  energy   consumpHon  from  a  staHc/dynamic  point  of  view   –  StaEc:  Finding  the  best  data  center  set-­‐up,  given  a   number  of  heterogeneous  machines   –  Dynamic:  OpHmizaHon  of  task  allocaHon  in  the   Resource  Manager   •  We  show  that  the  best  soluHon  implies  an   heterogeneous  data  center   –  Most  data  centers  are  heterogeneous  (several   generaHons  of  computers)   21   CCGrid 2012 José  M.Moya  |    Madrid  (Spain),  July  27,  2012   M.  Zapater,  J.M.  Moya,  J.L.  Ayala.  Leveraging  Heterogeneity  for   Energy  MinimizaHon  in  Data  Centers,  CCGrid  2012  
  22. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Current scenario 22

      WORKLOAD   Scheduler   Resource     Manager   ExecuEon   José  M.Moya  |    Madrid  (Spain),  July  27,  2012  
  23. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Potential improvements with

    best practices José  M.Moya  |    Madrid  (Spain),  July  27,  2012   23   ot aisle 0 200 400 600 800 1000 1200 1400 0 20 40 60 80 100 Power (KW) job size relative to data center capacity (%) Total power (computing and cooling) for various scheduling approaches savings by minimizing computing power savings by minimizing the recirculation’s effect savings by turning off idle machines unaddressed heat recirculation cost basic (unavoidable) cost max computing power, worst thermal placement min computing power, worst thermal placemenit optimal computing+cooling optimal computing+cooling, shut off idles optimal computing+cooling, shut off idles, no recirculation Fig. 3. Data center operation cost (in kilowatts) for various “savings modes”. Savings are based on heat recirculation data obtained by
  24. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Cooling-aware scheduling and

    resource allocation José  M.Moya  |    Madrid  (Spain),  July  27,  2012   24   0 50 100 150 200 FCFS-FF FCFS-LRH EDF-LRH FCFS-Xint SCINT energy consumed (GJ) Energy consumption, Scenario (a) 40 jobs, 25014 core-hours, idle servers on Throughput Turnaround time Alg. runtime Energy savings 0.197 jobs/hr 0.197 jobs/hr 0.172 jobs/hr 0.197 jobs/hr 0.163 jobs/hr 18.41 hr 18.41 hr 20.75 hr 18.41 hr 51.75 hr 3.4 ms 6.9 ms 213 ms 23 min 40 min 0% 6.2% 8.6% 8.7% 10.2% cooling energy computing energy (a) 0 5 10 15 20 25 30 35 40 FCFS-FF FCFS-LRH EDF-LRH FCFS-Xint SCINT energy consumed (GJ) Energy consumption, Scenario (a) 40 jobs, 25014 core-hours, idle servers off Throughput Turnaround time Alg. runtime Energy savings 0.197 jobs/hr 0.197 jobs/hr 0.172 jobs/hr 0.197 jobs/hr 0.163 jobs/hr 18.41 hr 18.41 hr 20.75 hr 18.41 hr 38.02 hr 3.4 ms 6.9 ms 213 ms 23 min 43 min 0% 11.8% 54.7% 21.8% 60.5% cooling energy computing energy (b) 0 50 100 150 200 250 300 FCFS-FF FCFS-LRH EDF-LRH FCFS-Xint SCINT energy consumed (GJ) Energy consumption, Scenario (b) 120 jobs, 16039 core-hours, idle servers on Throughput Turnaround time Alg. runtime Energy savings 0.580 jobs/hr 0.580 jobs/hr 0.349 jobs/hr 0.580 jobs/hr 0.254 jobs/hr 8.98 hr 8.98 hr 12.17 hr 8.98 hr 48.49 hr 170 ms 186 ms 397 ms 40.8 min 88.6 min 0% 1.7% 4.1% 3.6% 4.7% cooling energy computing energy (c) 0 5 10 15 20 25 30 35 40 FCFS-FF FCFS-LRH EDF-LRH FCFS-Xint SCINT energy consumed (GJ) Energy consumption, Scenario (b) 120 jobs, 16039 core-hours, idle servers off Throughput Turnaround time Alg. runtime Energy savings 0.580 jobs/hr 0.580 jobs/hr 0.349 jobs/hr 0.580 jobs/hr 0.427 jobs/hr 8.98 hr 8.98 hr 12.17 hr 8.98 hr 17.75 hr 171 ms 186 ms 397 ms 42 min 100 min 0% 4.0% 14.6% 14.2% 15.1% cooling energy computing energy (d) 300 350 400 450 (GJ) Energy consumption, Scenario (c) 174 jobs, 45817 core-hours, idle servers on Throughput Turnaround time Alg. runtime Energy savings 0.892 jobs/hr 0.892 jobs/hr 0.861 jobs/hr 0.892 jobs/hr 0.561 jobs/hr 9.99 hr 9.99 hr 13.39 hr 9.99 hr 65.38 hr 173 ms 196 ms 346 ms 20 min 142 min 0% 2.5% 5.9% 9.4% 12.5% cooling energy computing energy 80 100 (GJ) Energy consumption, Scenario (c) 174 jobs, 45817 core-hours, idle servers off Throughput Turnaround time Alg. runtime Energy savings 0.892 jobs/hr 0.892 jobs/hr 0.861 jobs/hr 0.892 jobs/hr 0.590 jobs/hr 9.99 hr 9.99 hr 13.39 hr 9.99 hr 61.49 hr 173 ms 191 ms 346 ms 21 min 147 min 0.0% 7.5% 17.3% 25.7% 41.4% cooling energy computing energy 0 50 100 150 FCFS-FF FCFS-LRH EDF-LRH FCFS-Xint SCINT energy consumed (GJ) Turnaround time Alg. runtime Energy savings 18.41 hr 18.41 hr 20.75 hr 18.41 hr 51.75 hr 3.4 ms 6.9 ms 213 ms 23 min 40 min 0% 6.2% 8.6% 8.7% 10.2% (a) 0 5 10 15 20 25 30 FCFS-FF FCFS-LRH EDF-LRH FCFS-Xint SCINT energy consumed (GJ) Throughput Turnaround time Alg. runtime Energy savings 0.197 jobs/hr 0.197 jobs/hr 0.172 jobs/hr 0.197 jobs/hr 0.163 jobs/hr 18.41 hr 18.41 hr 20.75 hr 18.41 hr 38.02 hr 3.4 ms 6.9 ms 213 ms 23 min 43 min 0% 11.8% 54.7% 21.8% 60.5% (b) 0 50 100 150 200 250 300 FCFS-FF FCFS-LRH EDF-LRH FCFS-Xint SCINT energy consumed (GJ) Energy consumption, Scenario (b) 120 jobs, 16039 core-hours, idle servers on Throughput Turnaround time Alg. runtime Energy savings 0.580 jobs/hr 0.580 jobs/hr 0.349 jobs/hr 0.580 jobs/hr 0.254 jobs/hr 8.98 hr 8.98 hr 12.17 hr 8.98 hr 48.49 hr 170 ms 186 ms 397 ms 40.8 min 88.6 min 0% 1.7% 4.1% 3.6% 4.7% cooling energy computing energy (c) 0 5 10 15 20 25 30 35 40 FCFS-FF FCFS-LRH EDF-LRH FCFS-Xint SCINT energy consumed (GJ) Energy consumption, Scenario (b) 120 jobs, 16039 core-hours, idle servers off Throughput Turnaround time Alg. runtime Energy savings 0.580 jobs/hr 0.580 jobs/hr 0.349 jobs/hr 0.580 jobs/hr 0.427 jobs/hr 8.98 hr 8.98 hr 12.17 hr 8.98 hr 17.75 hr 171 ms 186 ms 397 ms 42 min 100 min 0% 4.0% 14.6% 14.2% 15.1% cooling energy computing energy (d) 0 50 100 150 200 250 300 350 400 450 FCFS-FF FCFS-LRH EDF-LRH FCFS-Xint SCINT energy consumed (GJ) Energy consumption, Scenario (c) 174 jobs, 45817 core-hours, idle servers on Throughput Turnaround time Alg. runtime Energy savings 0.892 jobs/hr 0.892 jobs/hr 0.861 jobs/hr 0.892 jobs/hr 0.561 jobs/hr 9.99 hr 9.99 hr 13.39 hr 9.99 hr 65.38 hr 173 ms 196 ms 346 ms 20 min 142 min 0% 2.5% 5.9% 9.4% 12.5% cooling energy computing energy (e) 0 20 40 60 80 100 FCFS-FF FCFS-LRH EDF-LRH FCFS-Xint SCINT energy consumed (GJ) Energy consumption, Scenario (c) 174 jobs, 45817 core-hours, idle servers off Throughput Turnaround time Alg. runtime Energy savings 0.892 jobs/hr 0.892 jobs/hr 0.861 jobs/hr 0.892 jobs/hr 0.590 jobs/hr 9.99 hr 9.99 hr 13.39 hr 9.99 hr 61.49 hr 173 ms 191 ms 346 ms 21 min 147 min 0.0% 7.5% 17.3% 25.7% 41.4% cooling energy computing energy (f) Fig. 8. Energy comparison of the simulated schemes for the three scenarios. The plots correspond in respective positions to the plots of Figure 7. policy used in the data center, which enables job execution as soon as they arrive if the queue is empty and the data center is lightly loaded. In the “idle-on” case (Figure 8a), the total energy consumption using SCINT, EDF-LRH, iMPACT Lab (Arizona State U)
  25. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Application-aware scheduling and

    resource allocation 25   LSI-UPM WORKLOAD   Resource     Manager   (SLURM)   ExecuEon   Profiling  and   ClassificaEon   Energy     OpEmizaEon   José  M.Moya  |    Madrid  (Spain),  July  27,  2012  
  26. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Application-aware scheduling and

    resource allocation •  Workload:   –  12  tasks  from  SPEC  CPU  INT  2006   –  Random  workload  composed  by  2000  tasks,  divided  into   job  sets   –  Random  job  set  arrival  Hme   •  Servers:   26   Scenario José  M.Moya  |    Madrid  (Spain),  July  27,  2012  
  27. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Application-aware scheduling and

    resource allocation 27   Energy profiling WORKLOAD   Resource     Manager   (SLURM)   ExecuEon   Profiling  and   ClassificaEon   Energy     OpEmizaEon   Energy  profiling   José  M.Moya  |    Madrid  (Spain),  July  27,  2012  
  28. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Workload characterization José

     M.Moya  |    Madrid  (Spain),  July  27,  2012   28  
  29. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Application-aware scheduling and

    resource allocation 29   Optimization WORKLOAD   Resource     Manager   (SLURM)   ExecuEon   Profiling  and   ClassificaEon   Energy     OpEmizaEon   Energy  MinimizaEon:   •   MinimizaEon  subjected  to  constraints   •   MILP  problem  (solved  with  CPLEX)   •   StaEc  and  Dynamic   José  M.Moya  |    Madrid  (Spain),  July  27,  2012  
  30. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Application-aware scheduling and

    resource allocation 30   Static optimization •  DefiniHon  of  opHmal  data  center   –  Given  a  pool  of  100  servers  of  each  kind   –  1  job  set  from  workload   –  The  opHmizer  chooses  the  best  selecHon  of  servers   –  Constraints  of  cost  and  space   Best  soluEon:   •   40  Sparc   •   27  AMD     Savings:   •   5  a  22%  energy   •   30%  Eme   José  M.Moya  |    Madrid  (Spain),  July  27,  2012  
  31. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Application-aware scheduling and

    resource allocation 31   Dynamic optimization •  OpHmal  workload  allocaHon   –  Complete  workload  (2000  tasks)   –  Good  enough  resource  allocaHon  in  terms  of  energy  (not   the  best)   –  Run-­‐Hme  evaluaHon  and  opHmizaHon   Energy  savings   ranging  from  24%   to  47%   José  M.Moya  |    Madrid  (Spain),  July  27,  2012  
  32. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Application-aware scheduling and

    resource allocation •  First  proof-­‐of-­‐concept  regarding  the  use  of   heterogeneity  to  save  energy   •  AutomaHc  soluHon   •  AutomaHc  processor  selecHon  offers  notable  energy   savings   •  Easy  implementaHon  in  real  scenarios   –  SLURM  Resource  Manager   –  RealisHc  workloads  and  servers   32   Conclusions José  M.Moya  |    Madrid  (Spain),  July  27,  2012  
  33. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE 2. Server-level resource

    management José  M.Moya  |    Madrid  (Spain),  July  27,  2012   33   Chip   Server   Rack   Room   MulE-­‐ room   Sched  &  alloc   2 1   app   OS/middleware   Compiler/VM   3   3   architecture   4   4   technology   5  
  34. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Scheduling and resource

    allocation policies in MPSoCs A.  Coskun  ,  T.  Rosing  ,  K.  Whisnant  and  K.  Gross    "StaHc  and  dynamic  temperature-­‐ aware  scheduling  for  mulHprocessor  SoCs",    IEEE  Trans.  Very  Large  Scale  Integr.  Syst.,     vol.  16,    no.  9,    pp.1127  -­‐1140  2008     José  M.Moya  |    Madrid  (Spain),  July  27,  2012   34   Fig. 3. Distribution of thermal hot spots, with DPM (ILP). A. Static Scheduling Techniques We next provide an extensive comparison of the ILP based techniques. We refer to our static approach as Min-Th&Sp. As discussed in Section III, we implemented the ILP for min- imizing thermal hot spots (Min-Th), energy balancing (Bal- En), and energy minimization (Min-En) to compare against Fig. 4. Distribution of spatial gradients, with DPM (ILP). hot spots. While Min-Th reduces the high spatial differentials above 15 C, we observe a substantial increase in the spatial gradients above 10 C. In contrast, our method achieves lower and more balanced temperature distribution in the die. In Fig. 5, we show how the magnitudes of thermal cycles vary with the scheduling method. We demonstrate the average per- Fig. 3. Distribution of thermal hot spots, with DPM (ILP). A. Static Scheduling Techniques We next provide an extensive comparison of the ILP based techniques. We refer to our static approach as Min-Th&Sp. As discussed in Section III, we implemented the ILP for min- imizing thermal hot spots (Min-Th), energy balancing (Bal- Fig. 4. Distribution of spatial gradients, with DPM (ILP). hot spots. While Min-Th reduces the high spatial differentials above 15 C, we observe a substantial increase in the spatial gradients above 10 C. In contrast, our method achieves lower and more balanced temperature distribution in the die. In Fig. 5, we show how the magnitudes of thermal cycles vary UCSD – System Energy Efficiency Lab
  35. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Scheduling and resource

    allocation policies in MPSoCs •  Energy  characterizaHon  of  applicaHons  allows   to  define  proacHve  scheduling  and  resource   allocaHon  policies,  minimizing  hotspots   •  Hotspot  reducHon  allows  to  raise  cooling   temperature   +1oC  means  around  7%  cooling  energy  savings   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   35  
  36. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE 3. Application-aware and

    resource-aware virtual machine José  M.Moya  |    Madrid  (Spain),  July  27,  2012   36   Chip   Server   Rack   Room   MulE-­‐ room   Sched  &  alloc   2   1   app   OS/middleware   Compiler/VM   3 3 architecture   4   4   technology   5  
  37. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE JIT compilation in

    virtual machines •  Virtual  machines  compile   (JIT  compilaHon)  the   applicaHons  into  naHve   code  for  performance   reasons   •  The  opHmizer  is  general-­‐ purpose  and  focused  in   performance   opEmizaEon   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   37  
  38. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Back-­‐end   JIT

    compilation for energy minimization •  ApplicaHon-­‐aware  compiler   –  Energy  characterizaHon  of  applicaHons  and   transformaHons   –  ApplicaHon-­‐dependent  opHmizer   –  Global  view  of  the  data  center  workload   •  Energy  opHmizer   –  Currently,  compilers  for  high-­‐end  processors  oriented   to  performance  opHmizaHon   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   38   Front-­‐end   OpHmizer   Code  generator  
  39. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Energy saving potential

    for the compiler (MPSoCs) T.  Simunic,  G.  de  Micheli,  L.  Benini,  and  M.  Hans.  “Source  code  opHmizaHon  and   profiling  of  energy  consumpHon  in  embedded  systems,”  InternaHonal  Symposium  on   System  Synthesis,  pages  193  –  199,  Sept.  2000   – 77%  energy  reducHon  in  MP3  decoder   FEI,  Y.,  RAVI,  S.,  RAGHUNATHAN,  A.,  AND  JHA,  N.  K.  2004.  Energy-­‐opHmizing  source   code  transformaHons  for  OS-­‐driven  embedded  soyware.  In  Proceedings  of  the   InternaHonal  Conference  VLSI  Design.  261–266.   – Up  to  37,9%  (mean  23,8%)  energy  savings  in   mulHprocess  applicaHons  running  on  Linux   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   39  
  40. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE 4. Global automatic

    management of low-power modes José  M.Moya  |    Madrid  (Spain),  July  27,  2012   40   Chip   Server   Rack   Room   MulE-­‐ room   Sched  &  alloc   2   1   app   OS/middleware   Compiler/VM   3   3   architecture   4 4 technology   5  
  41. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE DVFS – Dynamic

    Voltage and Frequency Scaling •  As  supply  voltage  decreases,  power  decreases   quadraHcally   •  But  delay  increases  (performance  decreases)   only  linearly   •  The  maximum  frequency  also  decreases   linearly   •  Currently,  low-­‐power  modes,  if  used,  are   acHvated  by  inacHvity  of  the  server  operaHng   system   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   41  
  42. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Room-level DVFS • 

    To  minimize  energy  consumpHon,  changes   between  modes  should  be  minimized   •  There  exist  opHmal  algorithms  for  a  known   task  set  (YDS)   •  Workload  knowledge  allows  to  globally   schedule  low-­‐power  modes  without  any   impact  in  performance   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   42  
  43. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Parallelism to save

    energy Use of Parallelism Use of Parallelism V dd V dd /2 V dd /2 f f /2 f /2 f max f max /2 f max /2 9-17 Swiss Federal Institute of Technology Computer Engineering and Networks Laboratory José  M.Moya  |    Madrid  (Spain),  July  27,  2012   43  
  44. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE 5. Temperature-aware floorplanning

    of MPSoCs and many-cores José  M.Moya  |    Madrid  (Spain),  July  27,  2012   44   Chip   Server   Rack   Room   MulE-­‐ room   Sched  &  alloc   2   1   app   OS/middleware   Compiler/VM   3   architecture   4   4   technology   5
  45. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Temperature-aware floorplanning José

     M.Moya  |    Madrid  (Spain),  July  27,  2012   45  
  46. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Potential energy savings

    with floorplanning –  Up  to  21oC  reducHon  of  maximum  temperature   –  Mean:  -­‐12oC  in  maximum  temperature   –  Bejer  results  in  the  most  criHcal  examples   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   46   Temperature Reductions Average MaxTemp reduction: 12 oC Larger temperature reductions for benchmarks with higher maximum temperature For many benchmarks, temperature reducions are larger than 20 oC Maximum Temperature 0 20 40 60 80 100 120 140 ammp applu apsi art bzip2 crafty eon equake facerec fma3d gap gcc gzip lucas mcf mesa mgrid parser perlbmk swim twolf vortex vpr wupwise avg original modified Y.  Han,  I.  Koren,  and  C.  A.  Moritz.  Temperature  Aware  Floorplanning.  In  Proc.  of  the     Second  Workshop  on  Temperature-­‐Aware  Computer  Systems,  June  2005  
  47. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Temperature-aware floorplanning in

    3D chips José  M.Moya  |    Madrid  (Spain),  July  27,  2012   47   •  3D  chips  are  gezng  interest  due  to:   –  ↑  ↑    Scalability:  reduces  2D  equivalent   area   –  ↑  ↑    Performance:  shorter  wire  length   –  ↑  Reliability:  less  wiring     •  Drawback:   –  Huge  increment  of  hotspots       compared  with  2D  equivalent  designs  
  48. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Temperature-aware floorplanning in

    3D chips José  M.Moya  |    Madrid  (Spain),  July  27,  2012   48   •  Up  to  30oC  reducHon  per  layer  in  a  3D  chip   with  4  layers  and  48  cores  
  49. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE There is still

    much more to be done •  Smart  Grids   – Consume  energy  when  everybody  else  does  not   – Decrease  energy  consumpHon  when  everybody   else  is  consuming   •  Reducing  the  electricity  bill   – Variable  electricity  rates   – ReacHve  power  coefficient   – Peak  energy  demand   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   49  
  50. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Conclusions •  Reducing

     PUE  is  not  the  same  as  reducing  energy   consumpHon   –  IT  energy  consumpHon  dominates  in  state-­‐of-­‐the-­‐art  data   centers   •  ApplicaHon  and  resources  knowledge  can  be  effecHvely   used  to  define  proacEve  policies  to  reduce  the  total  energy   consumpHon   –  At  different  levels   –  In  different  scopes   –  Taking  into  account  cooling  and  computaHon  at  the  same  Hme   •  Proper  management  of  the  knowledge  of  the  data  center   thermal  behavior  can  reduce  reliability  issues   •  Reducing  energy  consumpHon  is  not  the  same  as  reducing   the  electricity  bill   José  M.Moya  |    Madrid  (Spain),  July  27,  2012   50  
  51. “Ingeniamos el futuro”! CAMPUS OF INTERNATIONAL EXCELLENCE Contact José  M.Moya

     |    Madrid  (Spain),  July  27,  2012   51   José M. Moya +34  607  082  892   jm.moya@upm.es     ETSI de Telecomunicación, B104 Avenida Complutense, 30 Madrid 28040, Spain Gracias: