1. GRASS: Trimming Stragglers in Approximation Analytics Ganesh Ananthanarayanan, Michael Hung, Xiaoqi Ren, Ion Stoica, Adam Wierman, Minlan Yu

2. Approximation Jobs Data deluge  timely answers from part of dataset are often good enough Deadline-bound Jobs: Maximize accuracy within a deadline – E.g., get me the best ad to show within 1s Error-bound Jobs: Minimize time taken to achieve desired accuracy – E.g., get me the #cars sold to the nearest 1000

3. Prioritize Tasks Spawn jobs on large dataset; subset of tasks need to complete – Accuracy α (Subset of completed tasks) Prioritize tasks – Jobs run only part of their tasks at a time (multi- waved) NP-Hard but many known heuristics…

4. Stragglers Occur despite prevention mechanisms Speculation: Spawn a duplicate, earliest wins Challenge: dynamically prioritize while choosing between speculative/unscheduled tasks

5. Speculative copies consume extra resources T3 Opportunity Cost T2 time Slot 1 Slot 2 Slot 3 T1 5 0 10 9 Is speculation worth the payoff? T1

6. Roadmap Two extreme approaches Switching between the two approaches Evaluation

7. Greedy Scheduling (GS) Schedule the task that greedily improves accuracy, i.e., finish earliest T1 T2 T3 time Slot 1 Slot 2 1 03 T4 T5 T6 6 T7 Task IDT1T2T3T4T5T6T7T8T9 t rem 5-------- t new 2---1111113 T1 Deadline = 6 Accuracy = 7/9

8. Resource Aware Scheduling (RAS) Speculate only if (c)t rem > (c+1)t new time T1 T2 T1 Slot 1 Slot 2 1 0 T6 T3 T4 T5 3 6 T7 T8 Task IDT1T2T3T4T5T6T7T8T9 t rem 5-------- t new 2---1111113 T1 Deadline = 6 Accuracy = 8/9

9. GS vs. RAS T1 T2 T3 time Slot 1 Slot 2 1 03 T4 T5 T6 6 T7 T1 Deadline = 6 Accuracy = 7/9 time T1 T2 T1 Slot 1 Slot 2 T6 T3 T4 T5 T7 T8 T1 Deadline = 6 Accuracy = 8/9 Deadline = 3 Accuracy = 3/9 T1 Accuracy = 2/9 GS RAS

10. How to pick between GS and RAS? When to consider opportunity cost? Analytical model [TBD]

11. Guideline Be “conservative” early in the job, and “aggressive” near its completion For jobs with >2 waves of tasks use RAS, otherwise use GS

12. GRASS = GS + RAS Start with RAS and switch to GS when 2 waves of tasks remain How to estimate two remaining waves? – Wave boundaries are not strict – Task durations are uncertain – Wave-width is not constant

13. Learning switching point in GRASS GRASS starts with RAS and switches to GS close to the deadline/error-bound Learn the switching point from completed jobs – Generate GS-only and RAS-only samples Switch when the GS-only samples are better for the remaining deadline

14. GRASS Scheduling GS  Greedy Scheduling RAS  Opportunity cost aware Scheduling Switch from RAS to GS close to deadline/error- bound Learn switching point

15. Implementation Details Implemented inside Hadoop 0.20.2 and Spark 0.7.3 – Modified Fair Scheduler Task Estimators – t rem is extrapolated from remaining data; progress reports at 5% intervals – t new is estimated from completed tasks

16. Evaluating GRASS Workload from Facebook and Bing traces – 1000’s of nodes, Hadoop and Dryad jobs – How was it generated? Baselines: LATE or Mantri 200 node EC2 deployment; trace-driven sim.

17. Deadline-bound Jobs

18. Error-bound Jobs

19. Value of Switching

20. Related Work

21. Conclusion