1 / 22

PBSE: A Robust Path-Based Speculative Execution for Data-Parallel Frameworks

PBSE: A Robust Path-Based Speculative Execution for Data-Parallel Frameworks. Riza O. Suminto , Cesar A. Stuardo , Alexandra Clark, Huan Ke , Tanakorn Leesatapornwongsa , Bo Fu, Daniar H. Kurniawan , Vincentius Martin, Uma Maheswara Rao G., Haryadi S. G unawi.

aarcher
Download Presentation

PBSE: A Robust Path-Based Speculative Execution for Data-Parallel Frameworks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. PBSE: A Robust Path-Based Speculative Execution for Data-Parallel Frameworks Riza O. Suminto, Cesar A. Stuardo, Alexandra Clark, HuanKe, TanakornLeesatapornwongsa, Bo Fu, Daniar H. Kurniawan, Vincentius Martin, Uma MaheswaraRao G., Haryadi S. Gunawi

  2. PBSE @ SoCC ’17 Speculative Execution (SE) 101 map task fast input data straggling slow NIC fastbackup task

  3. PBSE @ SoCC ’17 Is Hadoop always tail tolerant? Facebook Hadoopjobs on 30 nodes Normal (no slow NIC) vs. with 1slow 1Mbps NIC 1 job/hour!!! Why SE does not work?? 172 job/hour Many jobs cannot “escape” limping NIC

  4. PBSE @ SoCC ’17 Problem 1: No Straggler Detected • All slow  “No”straggler detected! slowNIC

  5. PBSE @ SoCC ’17 Problem 2: Straggling Backup • Backup task is also straggling! Many more cases fast slowNIC slow blame!

  6. PBSE @ SoCC ’17 Findings Basic Hadoop SE (BaseSE) • Good • Resource contention • Heterogeneous resources • Not Robust • Node-level network degradation • slow NIC and switches

  7. PBSE @ SoCC ’17 What’s The Flaw? • Network limpware is not considered a fault model • Task != Path • Slow paths are sometimes not exposed • Path progresses are lumped into per-task scores 4 paths 2 scores all Slow!

  8. PBSE @ SoCC ’17 Path-Based SE (PBSE) • Intuition: A taskis a collection of paths 1. PathDiversity 2. PathProgress 3. PathSpeculation • [Now] • Progresses • of 4paths: • M1R1 • M1R2 • M2R1 • M2R2 [Before] Progresses of 2tasks: R1 R2 1

  9. PBSE @ SoCC ’17 Contribution • Introduce node-level network degradation • Real important fault model • Expose path progress • Not just task progress • Develop pathstraggler detection & speculation • PBSE full integration to Hadoop • Initial integration to QFS, Spark, & Flume

  10. PBSE @ SoCC ’17 Outline • Intro • PBSE Techniques • Path Diversity • Path Progress • Path-based Speculation • Implementation & Evaluation • Conclusion

  11. PBSE @ SoCC ’17 Path Diversity • Enforce no potential SPOF-node • SPOF = single point of [tail-latency] failure • (to compare progresses of diverse paths)

  12. PBSE @ SoCC ’17 Path Progress • Each task reports path progresses/bandwidths • (not just report task scores) • Pathprogress can reveal the culprit • [Now] • 4 Paths: • M1R1 • M1R2 • M2R1 • M2R2 [Before] 2 Tasks: R1 R2 • [Now] • 4 Paths: • R1O1 • O1O1’ • R2O2 • O2O2’ [Before] 2 Tasks: R1 R2 No Straggler Actual straggler!

  13. PBSE @ SoCC ’17 Path-based Speculation • Use knowledge of previous failing paths • (don’t always blame the task/stage, blame the bottleneck)

  14. PBSE @ SoCC ’17 Outline • Intro • PBSE Techniques • Implementation & Evaluation • Conclusion

  15. PBSE @ SoCC ’17 Hadoop + PBSE • 6000+LOC over Hadoop/HDFS 2.7.1 • 3200 LOC in Application Manager • 1400 LOC in Task Management • 1400 LOC in HDFS

  16. PBSE @ SoCC ’17 PBSE vsBaseSE no tail 150 JobsFacebook Trace 15 nodes One 1-Mbps NIC speedup escape tail-SPOF

  17. PBSE @ SoCC ’17 Varied Experiment • Varying: • NIC degradations (60/30/10/1/0.1 Mbps) • Workload (Facebook, Cloudera) • Cluster size (15 to 60 nodes) • PBSE still gain speedups speedup-bw PBSE speedup at specific percentile of jobs

  18. PBSE @ SoCC ’17 PBSE vs Other Strategies • Others: • Vs. Other schedulers (Capacity / FIFO / Fair) • Vs. Other SE solutions (Cloning / Aggressive / HRead) • PBSE wins in all scenarios • We fix the fundamental flaws • i.e. path-based,not task-based PBSE Wins!

  19. PBSE @ SoCC ’17 Beyond MapReduce • All data-parallel systems need robust tail tolerance

  20. PBSE @ SoCC ’17 Outline • Intro • PBSE Techniques • Implementation & Evaluation • Conclusion

  21. PBSE @ SoCC ’17 Conclusion (Task abstraction) (Path abstraction) BaseSE + Resource Contention BaseSE + Limping NIC PBSE + Limping NIC long tail latency! can not escape tail More Robust SE

  22. SystemName @ ConfName ’14 Thank you!Questions? http://ucare.cs.uchicago.edu

More Related