Successfully reported this slideshow.
Your SlideShare is downloading.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.
Energy Efficient Scheduling for High-Performance Clusters<br />ZiliangZong, Texas State University <br />Adam Manzanares, ...
Where is Auburn University?<br />Ph.D.’04, U. of Nebraska-Lincoln<br />04-07, New Mexico Tech<br />07-now, Auburn Universi...
Storage Systems Research Group at New Mexico Tech (2004-2007)<br />2011/6/22<br />3<br />
Storage Systems Research Group at Auburn (2008)<br />2011/6/22<br />4<br />
Storage Systems Research Group at Auburn (2009)<br />2011/6/22<br />5<br />
Storage Systems Research Group at Auburn (2011)<br />2011/6/22<br />6<br />
Investigators<br />ZiliangZong, Ph.D. <br />	Assistant Professor, <br />   Texas State University<br />Adam Manzanares, Ph...
2011/6/22<br />8<br />Introduction - Applications<br />
Introduction – Data Centers<br />2011/6/22<br />9<br />
Motivation – Electricity Usage<br />EPA Report to Congress on Server and Data Center Energy Efficiency, 2007<br />2011/6/2...
Motivation – Energy Projections<br />EPA Report to Congress on Server and Data Center Energy Efficiency, 2007<br />2011/6/...
Motivation – Design Issues<br />2011/6/22<br />12<br />
Architecture – Multiple Layers<br />2011/6/22<br />13<br />
Energy Efficient Devices<br />2011/6/22<br />14<br />
Multiple Design Goals<br />2011/6/22<br />15<br />
Energy-Aware Scheduling for Clusters<br />2011/6/22<br />16<br />
Parallel Applications<br />2011/6/22<br />17<br />
Motivational Example<br />8<br />T1<br />T3<br />T2<br />T4<br />1<br />23<br />33<br />39<br />0<br />8<br />6<br />5<br ...
Motivational Example (cont.)<br />(8,48)<br />(6,6)<br />(5,5)<br />T1<br />T3<br />T2<br />T4<br />1<br />23<br />33<br /...
Motivational Example (cont.)<br />(8,48)<br />(6,6)<br />(5,5)<br />1<br />(15,90)<br />(10,60)<br />2<br />3<br />T1<br /...
Basic Steps of Energy-Aware Scheduling<br />Algorithm Implementation:<br />Step 1: DAG Generation<br />Task Description:<b...
Basic Steps of Energy-Aware Scheduling<br />Algorithm Implementation:<br />Total Execution time from current task to the e...
Basic Steps of Energy-Aware Scheduling<br />Algorithm Implementation:<br />Original Task List: {10, 9, 8, 5, 6, 2, 7, 4, 3...
Basic Steps of Energy-Aware Scheduling<br />Algorithm Implementation:<br />Original Task List: {10, 9, 8, 5, 6, 2, 7, 4, 3...
The EAD and PEBD Algorithms<br />Generate the DAG of given task sets<br />Calculate energy increase<br />and time decrease...
Energy Dissipation in Processors<br />http://www.xbitlabs.com<br />2011/6/22<br />26<br />
Parallel Scientific Applications<br />Fast Fourier Transform<br />Gaussian Elimination<br />2011/6/22<br />27<br />
Large-Scale Parallel Applications <br />Robot Control<br />Sparse Matrix Solver<br />http://www.kasahara.elec.waseda.ac.jp...
Impact of CPU Power Dissipation<br />Impact of CPU Types:<br />19.4%<br />3.7%<br />Energy consumption for different proce...
Impact of Interconnect Power Dissipation<br />Impact of Interconnection Types:<br />5%<br />3.1%<br />16.7%<br />13.3%<br ...
Parallelism Degrees<br />Impact of Application Parallelism:<br />6.9%<br />5.4%<br />17%<br />15.8%<br />Energy consumptio...
Communication-Computation Ratio<br />Impact of CCR:<br />Energy consumption under different CCRs<br />CCR: Communication-C...
Performance<br />Impact to Schedule Length:<br />Schedule length of Gaussian Elimination<br />Schedule length of Sparse Ma...
Heterogeneous Clusters - Motivational Example<br />2011/6/22<br />34<br />
Motivational Example (cont.)<br />Energy calculation for tentative schedule<br />C1<br />C2<br />C3<br />C4<br />2011/6/22...
Experimental Settings<br />Simulation Environments<br />2011/6/22<br />36<br />
Communication-Computation Ratio<br />CCR sensitivity for Gaussian Elimination<br />2011/6/22<br />37<br />
Heterogeneity<br />Computational nodes heterogeneity experiments<br />2011/6/22<br />38<br />
Conclusions<br /><ul><li>Architecture for high-performance computing platforms
Energy-Efficient Scheduling for Clusters
Energy-Efficient Scheduling for Heterogeneous Systems
How to measure energy consumption? Kill-A-Watt</li></ul>2011/6/22<br />39<br />
Source Code Availability<br />www.mcs.sdsmt.edu/~zzong/software/scheduling.html<br />2011/6/22<br />40<br />
Upcoming SlideShare
Loading in …5
×

1

Share

Download Now Download

Download to read offline

Energy efficient resource management for high-performance clusters

Download Now Download

Download to read offline

In the past decade, high-performance cluster computing platforms have been widely used to solve challenging and rigorous engineering tasks in industry and scientific applications. Due to extremely high energy cost,reducing energy consumption has become a major
concern in designing economical and environmentally friendly cluster computing
infrastructures for many high-performance applications. The primary focus of this talk is to illustrate how to improve energy efficiency of clusters and storage systems without significantly degrading performance. In this talk, we will first describe a general architecture
for building energy-efficient cluster computing platforms. Then, we will outline several energyefficient scheduling algorithms designed for high-performance clusters and large-scale storage systems. The experimental results using both synthetic and real world applications
show that energy dissipation in clusters can be reduced with a marginal degradation of system performance.

Related Books

Free with a 30 day trial from Scribd

See all

Related Audiobooks

Free with a 30 day trial from Scribd

See all
  • Be the first to comment

Energy efficient resource management for high-performance clusters

  1. 1. Energy Efficient Scheduling for High-Performance Clusters<br />ZiliangZong, Texas State University <br />Adam Manzanares, Los Alamos National Lab <br />Xiao Qin, Auburn University<br />
  2. 2. Where is Auburn University?<br />Ph.D.’04, U. of Nebraska-Lincoln<br />04-07, New Mexico Tech<br />07-now, Auburn University<br />
  3. 3. Storage Systems Research Group at New Mexico Tech (2004-2007)<br />2011/6/22<br />3<br />
  4. 4. Storage Systems Research Group at Auburn (2008)<br />2011/6/22<br />4<br />
  5. 5. Storage Systems Research Group at Auburn (2009)<br />2011/6/22<br />5<br />
  6. 6. Storage Systems Research Group at Auburn (2011)<br />2011/6/22<br />6<br />
  7. 7. Investigators<br />ZiliangZong, Ph.D. <br /> Assistant Professor, <br /> Texas State University<br />Adam Manzanares, Ph.D. Candidate <br />Los Alamos National Lab<br />Xiao Qin, Ph.D. <br />Associate Professor <br /> Auburn University<br />2011/6/22<br />7<br />
  8. 8. 2011/6/22<br />8<br />Introduction - Applications<br />
  9. 9. Introduction – Data Centers<br />2011/6/22<br />9<br />
  10. 10. Motivation – Electricity Usage<br />EPA Report to Congress on Server and Data Center Energy Efficiency, 2007<br />2011/6/22<br />10<br />
  11. 11. Motivation – Energy Projections<br />EPA Report to Congress on Server and Data Center Energy Efficiency, 2007<br />2011/6/22<br />11<br />
  12. 12. Motivation – Design Issues<br />2011/6/22<br />12<br />
  13. 13. Architecture – Multiple Layers<br />2011/6/22<br />13<br />
  14. 14. Energy Efficient Devices<br />2011/6/22<br />14<br />
  15. 15. Multiple Design Goals<br />2011/6/22<br />15<br />
  16. 16. Energy-Aware Scheduling for Clusters<br />2011/6/22<br />16<br />
  17. 17. Parallel Applications<br />2011/6/22<br />17<br />
  18. 18. Motivational Example<br />8<br />T1<br />T3<br />T2<br />T4<br />1<br />23<br />33<br />39<br />0<br />8<br />6<br />5<br />2<br />3<br />T1<br />T3<br />T4<br />10<br />15<br />23<br />26<br />32<br />0<br />8<br />6<br />2<br />2<br />4<br />T2<br />4<br />24<br />14<br />6<br />T3<br />T4<br />T1<br />T1<br />23<br />29<br />20<br />0<br />8<br />0<br />8<br />2<br />T2<br />18<br />Linear Schedule<br />Time: 39s<br />No Duplication Schedule (NDS)<br />Time: 32s<br />Task Duplication Schedule (TDS)<br />Time: 29s<br />An Example of duplication<br />2011/6/22<br />18<br />
  19. 19. Motivational Example (cont.)<br />(8,48)<br />(6,6)<br />(5,5)<br />T1<br />T3<br />T2<br />T4<br />1<br />23<br />33<br />39<br />0<br />8<br />(15,90)<br />(10,60)<br />2<br />3<br />T1<br />T3<br />T4<br />(4,4)<br />(2,2)<br />23<br />26<br />32<br />0<br />8<br />6<br />2<br />T2<br />(6,36)<br />4<br />24<br />14<br />T3<br />T4<br />T1<br />T1<br />23<br />29<br />20<br />0<br />8<br />0<br />8<br />2<br />T2<br />18<br />Linear Schedule<br />Time:39s Energy: 234J <br />No Duplication Schedule (MCP)<br />Time: 32s Energy: 242J<br />Task Duplication Schedule (TDS)<br />Time: 29s Energy: 284J<br />An Example of duplication<br />CPU_Energy=6W<br />Network_Energy=1W<br />2011/6/22<br />19<br />
  20. 20. Motivational Example (cont.)<br />(8,48)<br />(6,6)<br />(5,5)<br />1<br />(15,90)<br />(10,60)<br />2<br />3<br />T1<br />T3<br />T4<br />(4,4)<br />(2,2)<br />23<br />26<br />32<br />0<br />8<br />6<br />2<br />T2<br />(6,36)<br />4<br />24<br />14<br />T3<br />T4<br />T1<br />T1<br />23<br />29<br />20<br />0<br />8<br />0<br />8<br />2<br />T2<br />18<br />The energy cost of duplicating T1:<br />CPU side: 48J Network side: -6J Total: 42J<br />The performance benefit of duplicating T1: 6s<br />Energy-performance tradeoff: 42/6 = 7<br />EAD<br />Time: 32s Energy: 242J<br />PEBD<br />Time: 29s Energy: 284J<br />If Threshold = 10 <br />Duplicate T1? <br />EAD: NO <br />PEBD: Yes<br />2011/6/22<br />20<br />
  21. 21. Basic Steps of Energy-Aware Scheduling<br />Algorithm Implementation:<br />Step 1: DAG Generation<br />Task Description:<br />Task Set {T1, T2, …, T9, T10 }<br />T1 is the entry task;<br />T10 is the exit task;<br />T2, T3 and T4 can not start until T1 finished;<br />T5 and T6 can not start until T2 finished;<br />T7 can not start until both T3 and T4 finished;<br />T8 can not start until both T5 and T6 finished;<br />T9 can not start until both T6 and T7 finished;<br />T10 can not start until both T8 and T9 finished;<br />2011/6/22<br />21<br />
  22. 22. Basic Steps of Energy-Aware Scheduling<br />Algorithm Implementation:<br />Total Execution time from current task to the exit task<br />Earliest Start Time<br />Earliest Completion Time<br />Latest Allowable Start Time<br />Latest Allowable Completion Time<br />Favorite Predecessor<br />Step 2: Parameters Calculation<br />2011/6/22<br />22<br />
  23. 23. Basic Steps of Energy-Aware Scheduling<br />Algorithm Implementation:<br />Original Task List: {10, 9, 8, 5, 6, 2, 7, 4, 3, 1} <br />Original Task List: {10, 9, 8, 5, 6, 2, 7, 4, 3,1} <br />Original Task List: {10, 9, 8, 5, 6, 2, 7, 4, 3,1} <br />Original Task List: {10, 9, 8,5, 6, 2, 7, 4, 3,1} <br />Original Task List: {10, 9, 8,5, 6, 2, 7,4, 3,1} <br />Step 3: Scheduling<br />2011/6/22<br />23<br />
  24. 24. Basic Steps of Energy-Aware Scheduling<br />Algorithm Implementation:<br />Original Task List: {10, 9, 8, 5, 6, 2, 7, 4, 3, 1} <br />Original Task List: {10, 9, 8, 5, 6, 2, 7, 4, 3,1} <br />Original Task List: {10, 9, 8, 5, 6, 2, 7, 4, 3,1} <br />Original Task List: {10, 9, 8,5, 6, 2, 7, 4, 3,1} <br />Original Task List: {10, 9, 8,5, 6, 2, 7,4, 3,1} <br />Step 4: Duplication Decision<br />Decision 1: Duplicate T1?<br />Decision 2: Duplicate T2?<br /> Duplicate T1?<br />Decision 3: Duplicate T1?<br />2011/6/22<br />24<br />
  25. 25. The EAD and PEBD Algorithms<br />Generate the DAG of given task sets<br />Calculate energy increase<br />and time decrease<br />Calculate energy increase<br />Find all the critical paths in DAG<br />Ratio= energy increase/ time decrease<br />more_energy<=Threshold?<br />Generate scheduling queue based on the level (ascending)<br />No<br />Yes<br />select the task (has not been scheduled yet) with the lowest level as starting task <br />No<br />Ratio<=Threshold?<br />Duplicate this task and select the next task in the same critical path<br />Yes<br />meet entry task<br />Duplicate this task and select the next task in the same critical path<br />No<br />allocate it to the same processor with the tasks in the same critical path<br />Yes<br />No<br />For each task which is in the <br />same critical path with starting task, check<br /> if it is already scheduled <br />Save time if duplicate <br />this task?<br />Yes<br />PEBD<br />EAD<br />2011/6/22<br />25<br />
  26. 26. Energy Dissipation in Processors<br />http://www.xbitlabs.com<br />2011/6/22<br />26<br />
  27. 27. Parallel Scientific Applications<br />Fast Fourier Transform<br />Gaussian Elimination<br />2011/6/22<br />27<br />
  28. 28. Large-Scale Parallel Applications <br />Robot Control<br />Sparse Matrix Solver<br />http://www.kasahara.elec.waseda.ac.jp/schedule/<br />2011/6/22<br />28<br />
  29. 29. Impact of CPU Power Dissipation<br />Impact of CPU Types:<br />19.4%<br />3.7%<br />Energy consumption for different processors (Gaussian, CCR=0.4) <br />Energy consumption for different processors (FFT, CCR=0.4) <br />2011/6/22<br />29<br />
  30. 30. Impact of Interconnect Power Dissipation<br />Impact of Interconnection Types:<br />5%<br />3.1%<br />16.7%<br />13.3%<br />Energy consumption (Robot Control, Myrinet) <br />Energy consumption (Robot Control, Infiniband) <br />2011/6/22<br />30<br />
  31. 31. Parallelism Degrees<br />Impact of Application Parallelism:<br />6.9%<br />5.4%<br />17%<br />15.8%<br />Energy consumption of Sparse Matrix (Myrinet)<br />Energy consumption of Robert Control(Myrinet)<br />2011/6/22<br />31<br />
  32. 32. Communication-Computation Ratio<br />Impact of CCR:<br />Energy consumption under different CCRs<br />CCR: Communication-Computation Rate<br />2011/6/22<br />32<br />
  33. 33. Performance<br />Impact to Schedule Length:<br />Schedule length of Gaussian Elimination<br />Schedule length of Sparse Matrix Solver<br />2011/6/22<br />33<br />
  34. 34. Heterogeneous Clusters - Motivational Example<br />2011/6/22<br />34<br />
  35. 35. Motivational Example (cont.)<br />Energy calculation for tentative schedule<br />C1<br />C2<br />C3<br />C4<br />2011/6/22<br />35<br />
  36. 36. Experimental Settings<br />Simulation Environments<br />2011/6/22<br />36<br />
  37. 37. Communication-Computation Ratio<br />CCR sensitivity for Gaussian Elimination<br />2011/6/22<br />37<br />
  38. 38. Heterogeneity<br />Computational nodes heterogeneity experiments<br />2011/6/22<br />38<br />
  39. 39. Conclusions<br /><ul><li>Architecture for high-performance computing platforms
  40. 40. Energy-Efficient Scheduling for Clusters
  41. 41. Energy-Efficient Scheduling for Heterogeneous Systems
  42. 42. How to measure energy consumption? Kill-A-Watt</li></ul>2011/6/22<br />39<br />
  43. 43. Source Code Availability<br />www.mcs.sdsmt.edu/~zzong/software/scheduling.html<br />2011/6/22<br />40<br />
  44. 44. Download the presentation slideshttp://www.slideshare.net/xqin74<br />Google: slideshare Xiao Qin<br />‹#› <br />
  45. 45. http://www.eng.auburn.edu/~xqin<br />
  46. 46. My webpagehttp://www.eng.auburn.edu/~xqin<br />
  47. 47. Download Slides at slidesharehttp://www.slideshare.net/xqin74<br />
  48. 48. Questionshttp://www.eng.auburn.edu/~xqin<br />2011/6/22<br />45<br />

    Be the first to comment

    Login to see the comments

  • AishaShabbir1

    Mar. 7, 2017

In the past decade, high-performance cluster computing platforms have been widely used to solve challenging and rigorous engineering tasks in industry and scientific applications. Due to extremely high energy cost,reducing energy consumption has become a major concern in designing economical and environmentally friendly cluster computing infrastructures for many high-performance applications. The primary focus of this talk is to illustrate how to improve energy efficiency of clusters and storage systems without significantly degrading performance. In this talk, we will first describe a general architecture for building energy-efficient cluster computing platforms. Then, we will outline several energyefficient scheduling algorithms designed for high-performance clusters and large-scale storage systems. The experimental results using both synthetic and real world applications show that energy dissipation in clusters can be reduced with a marginal degradation of system performance.

Views

Total views

1,484

On Slideshare

0

From embeds

0

Number of embeds

2

Actions

Downloads

46

Shares

0

Comments

0

Likes

1

×