power aware job scheduling
play

POWER-AWARE JOB SCHEDULING Maximizing Data Center Performance - PowerPoint PPT Presentation

POWER-AWARE JOB SCHEDULING Maximizing Data Center Performance Under Strict Power Budget Osman Sarood, Akhil Langer , Abhishek Gupta, Laxmikant Kale Parallel


  1. POWER-­‑AWARE ¡JOB ¡SCHEDULING ¡ Maximizing ¡Data ¡Center ¡Performance ¡Under ¡Strict ¡Power ¡Budget ¡ Osman ¡Sarood, ¡ Akhil ¡Langer , ¡Abhishek ¡Gupta, ¡Laxmikant ¡Kale ¡ ¡ Parallel ¡Programming ¡Laboratory ¡ Department ¡of ¡Computer ¡Science ¡ University ¡of ¡Illinois ¡at ¡Urbana-­‑Champaign ¡ ¡ 29 th ¡April ¡2014 ¡ ¡ ¡

  2. Major ¡Challenges ¡to ¡Achieve ¡Exascale 1 ¡ q Energy ¡and ¡Power ¡Challenge ¡ q Memory ¡and ¡Storage ¡Challenge ¡ q Concurrency ¡and ¡Locality ¡Challenge ¡ q Resiliency ¡Challenge ¡ Kogge, ¡Peter, ¡et ¡al. ¡"Exascale ¡compuWng ¡study: ¡Technology ¡challenges ¡in ¡achieving ¡exascale ¡systems." ¡(2008). ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 2 ¡

  3. Major ¡Challenges ¡to ¡Achieve ¡Exascale 1 ¡ Power ¡consumpWon ¡for ¡Top500 ¡ Exascale ¡in ¡ 20MW! ¡ Kogge, ¡Peter, ¡et ¡al. ¡"Exascale ¡compuWng ¡study: ¡Technology ¡challenges ¡in ¡achieving ¡exascale ¡systems." ¡(2008). ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 3 ¡

  4. Data ¡Center ¡Power ¡ How ¡is ¡data ¡center ¡power ¡need ¡calculated? ¡ ¡ q using ¡Thermal ¡Design ¡Power ¡(TDP) ¡of ¡nodes ¡ However, ¡TDP ¡is ¡hardly ¡reached!! ¡ ¡ ¡ SoluWon ¡ q constrain ¡power ¡consumpWon ¡of ¡nodes ¡ q Overprovisioning ¡-­‑ ¡Use ¡more ¡nodes ¡than ¡convenWonal ¡ data ¡center ¡for ¡the ¡same ¡power ¡budget ¡ ¡ ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 4 ¡ ¡

  5. DistribuWon ¡of ¡Node ¡Power ¡ConsumpWon ¡ Power ¡distribuWon ¡for ¡BG/Q ¡ processor ¡on ¡Mira ¡ q 76% ¡by ¡CPU/Memory ¡ q No ¡good ¡mechanism ¡for ¡ controlling ¡other ¡power ¡ domains ¡ ¡ ¡ ¡ Pie ¡Chart: ¡Sean ¡Wallace, ¡Measuring ¡Power ¡ConsumpWon ¡on ¡IBM ¡Blue ¡Gene/Q ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 5 ¡

  6. Constraining ¡CPU/Memory ¡Power ¡ Intel ¡Sandy ¡Bridge ¡ q Running ¡Average ¡Power ¡Limit ¡(RAPL) ¡library ¡ Ø measure ¡and ¡set ¡CPU/memory ¡power ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 6 ¡

  7. ApplicaWon ¡Performance ¡with ¡Power ¡ Configura7on ¡ ¡ p c : ¡CPU ¡power ¡cap ¡ P m :Memory ¡power ¡cap ¡ (n ¡x ¡p c , ¡p m ¡) ¡ q ApplicaWon ¡performance ¡does ¡not ¡ improve ¡proporWonately ¡with ¡ increase ¡in ¡power ¡cap ¡ q Beher ¡is ¡to ¡run ¡on ¡larger ¡number ¡ of ¡nodes ¡each ¡capped ¡at ¡lower ¡ power ¡level ¡ (12x44,18) ¡ ¡ (20x32,10) ¡ ¡ Performance ¡of ¡LULESH ¡at ¡different ¡configuraWons ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 7 ¡

  8. Problem ¡Statement ¡ ¡ ¡ Maximizing ¡Data ¡Center ¡Performance ¡Under ¡ ¡Strict ¡ Power ¡Budget ¡ ¡ Data ¡center ¡capabiliWes ¡and ¡job ¡features ¡ q Power ¡capping ¡ability ¡ q Overprovisioning ¡ q Moldability ¡(OpWonal) ¡ q Malleability ¡(OpWonal) ¡ Ø Charm++ ¡ Ø Dynamic ¡MPI ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 8 ¡

  9. Power ¡Aware ¡Resource ¡Manager ¡(PARM) ¡ JOB ¡PROFILER ¡ EXECUTION ¡ SCHEDULER ¡ PASS ¡MODEL ¡ FRAMEWORK ¡ ¡ q SHRINK/EXPAND ¡JOBS ¡ q APPLY ¡POWER ¡CAPS ¡ ¡ JOB ¡QUEUE ¡ JOB ¡ TRIGGERS ¡ JOB ¡ARRIVAL ¡ TERMINATION ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 9 ¡

  10. JOB ¡PROFILER ¡ q Measure ¡job ¡performance ¡at ¡various ¡scales ¡ and ¡cpu ¡power ¡caps ¡ q Power ¡Aware ¡Strong ¡Scaling ¡(PASS) ¡Model ¡ ¡ Ø Predict ¡job ¡performance ¡at ¡any ¡(n, ¡p) ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 10 ¡

  11. Power ¡Aware ¡Strong ¡Scaling ¡(PASS) ¡Model ¡ Time ¡vs ¡Scale ¡ Downey’s ¡strong ¡scaling ¡ q p core : ¡core ¡power ¡ Time ¡vs ¡Frequency ¡ t = F ( n , A , σ ) q g i : ¡cost ¡level ¡I ¡cache ¡access ¡ q L i : ¡#level ¡I ¡accesses ¡ 8 W cpu + T mem , for f < f h q n: ¡number ¡of ¡nodes ¡ < q g m : ¡cost ¡of ¡mem ¡access ¡ t ( f ) = f q A: ¡Average ¡Parallelism ¡ q M: ¡#mem ¡accesses ¡ T h , for f ≥ f h : q ¡σ ¡: ¡duraWon ¡of ¡parallelism ¡A ¡ q p base : ¡idle ¡power ¡ Frequency ¡vs ¡Power ¡ q W cpu : ¡CPU ¡work ¡ xpressed as [31]: q T mem : ¡memory ¡work ¡ 3 X p = p core + g i L i + g m M + p base q T h ¡ : ¡ ¡ ¡minimum ¡exec ¡Wme ¡ ¡ i =1 is the base/static package power consumption. Time ¡as ¡a ¡funcDon ¡of ¡power ¡and ¡number ¡of ¡nodes ¡ ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 11 ¡

  12. Power ¡Aware ¡Resource ¡Manager ¡(PARM) ¡ JOB ¡PROFILER ¡ EXECUTION ¡ SCHEDULER ¡ PASS ¡MODEL ¡ FRAMEWORK ¡ ¡ q SHRINK/EXPAND ¡JOBS ¡ q APPLY ¡POWER ¡CAPS ¡ ¡ JOB ¡QUEUE ¡ JOB ¡ TRIGGERS ¡ JOB ¡ARRIVAL ¡ TERMINATION ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 12 ¡

  13. Scheduler: ¡Integer ¡Linear ¡Program ¡FormulaWon ¡ Objective Function X X X w j ∗ s j,n,p ∗ x j,n,p j ∈ J n ∈ N j p ∈ P j Select One Resource Combination Per Job X X x j,n,p ≤ 1 ∀ j ∈ I n ∈ N j p ∈ P j X X x j,n,p = 1 ∀ j ∈ I n ∈ N j p ∈ P j Bounding total nodes X X X nx j,n,p ≤ N j ∈ J p ∈ P j n ∈ N j Bounding power consumption X X X ( n ∗ ( p + W base )) x j,n,p ≤ W max j ∈ J n ∈ N j p ∈ P j Disable Malleability (Optional) X X nx j,n,p = n j ∀ j ∈ I n ∈ N j p ∈ P j 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 13 ¡

  14. Scheduler: ¡ObjecWve ¡FuncWon ¡ q Maximizing ¡throughput ¡makes ¡ILP ¡opWmizaWon ¡infeasible ¡ q Maximize ¡sum ¡of ¡power-­‑aware ¡speedup ¡of ¡selected ¡jobs: ¡ s j,n,p = t j,min ( N j ) ,min ( P j ) t j,n,p is the speedup of job executing 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 14 ¡

  15. Power ¡Aware ¡Resource ¡Manager ¡(PARM) ¡ JOB ¡PROFILER ¡ EXECUTION ¡ SCHEDULER ¡ PASS ¡MODEL ¡ FRAMEWORK ¡ ¡ q SHRINK/EXPAND ¡JOBS ¡ q APPLY ¡POWER ¡CAPS ¡ ¡ JOB ¡QUEUE ¡ JOB ¡ TRIGGERS ¡ JOB ¡ARRIVAL ¡ TERMINATION ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 15 ¡

  16. Experimental ¡Setup ¡ q ApplicaWons ¡ q Job ¡Dataset ¡ Ø Memory-­‑intensive ¡ q ¡β ¡corresponds ¡to ¡CPU ¡sensiWvity ¡ § Jacobi ¡and ¡Wave2D ¡ q SetL: ¡Mix ¡of ¡apps ¡with ¡average ¡β=0.1 ¡ Ø ComputaWon-­‑intensive ¡ q SetH: ¡Mix ¡of ¡apps ¡with ¡average ¡β=0.27 ¡ § LeanMD ¡ Ø Mixed ¡ § AMR ¡and ¡Lulesh ¡ q Power ¡Budget ¡ q Testbed ¡ q 38-­‑node ¡Intel ¡Sandy ¡Bridge ¡ q CPU ¡power ¡levels={30, ¡32, ¡34, ¡39, ¡45, ¡55}W ¡ q 6 ¡physical ¡cores, ¡16GB ¡RAM ¡ ¡ q Node ¡power ¡consumpWon= ¡116W ¡ q Power ¡capping ¡using ¡RAPL ¡ q Power ¡Budget ¡= ¡3000W ¡ q CPU ¡power ¡cap ¡range ¡[25-­‑95]W ¡ q #nodes ¡in ¡tradiWonal ¡data ¡center ¡= ¡28 ¡ ¡ 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 16 ¡

  17. EsWmaWng ¡Performance ¡using ¡PASS ¡ 1.8 LeanMD AMR 1.7 Model ¡Parameters ¡ Lulesh Wave2D 1.6 Power − aware speedup Jacobi2D 1.5 1.4 1.3 1.2 1.1 1 30 35 40 45 50 55 60 CPU power (W) 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 17 ¡

  18. PARM ¡Performance ¡Results ¡ Average ¡CompleWon ¡Wmes ¡ 300 Descrip(on ¡ SLURM noMM q noMM : ¡without ¡Malleability ¡and ¡Moldability ¡ Average completion time (mins) 250 noSE wSE q noSE : ¡ ¡ ¡ ¡with ¡Moldability ¡but ¡no ¡Malleability ¡ 200 q wSE : ¡ ¡ ¡ ¡ ¡ ¡with ¡Moldability ¡and ¡Malleability ¡ 150 Performance ¡ q 32% ¡improvement ¡with ¡nMM ¡over ¡SLURM ¡ 100 q 13.9% ¡improvement ¡with ¡noSE ¡over ¡noMM ¡ 50 q 7.5% ¡improvement ¡with ¡wSE ¡over ¡noSE ¡ q 1.7X ¡improvement ¡in ¡throughput ¡ 0 SetL SetH 3/4/15 ¡ Power-­‑Aware ¡Job ¡Scheduling ¡ 18 ¡

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend


More recommend