1. Improving MapReduce Performance Using Smart Speculative Execution Strategy Qi Chen, Cheng Liu, and Zhen Xiao Oct 2013 To appear in IEEE Transactions on Computers

2. 0 Outlines 2. Background 3. Previous work 4. Pitfalls 5. Our Design 1. Introduction 6. Evaluation 7. Conclusion

3. 0 Outlines 2. Background 3. Previous work 4. Pitfalls 5. Our Design 1. Introduction 6. Evaluation 7. Conclusion

4. Introduction  The new era of Big Data is coming!  – 20 PB per day (2008)  – 30 TB per day (2009)  – 60 TB per day (2010)  –petabytes per day  What does big data mean?  Important user information  significant business value

5. MapReduce  What is MapReduce?  most popular parallel computing model proposed by Google database operation Search engine Machine learning Cryptanalysi s Scientific computation Applications … Select, Join, Group Page rank, Inverted index, Log analysis Clustering, machine translation, Recommendation

6. Straggler  What is straggler in MapReduce?  Nodes on which tasks take an unusually long time to finish  It will:  Delay the job execution time  Degrade the cluster throughput  How to solve it?  Speculative execution  Slow task is backed up on an alternative machine with the hope that the backup one can finish faster

7. 0 Outlines 2. Background 3. Previous work 4. Pitfalls 5. Our Design 1. Introduction 6. Evaluation 7. Conclusion

8. Architecture Split 1 Split 2 … Split M Map Part 2 Part 1 Map Part 2 Part 1 Map Part 2 Part 1 Reduc e Output2 Input files Map Stage Reduce Stage Output files Output1 Master … Assig n

9. Programming model  Input : (key, value) pairs  Output : (key*, value*) pairs Phase Stage Map: MapCombine List(K 1, V 1 ) → List(K 2,V 2 ) → List(K 2, List(V 2 )) Reduce: CopySortReduce List(K 2, List(V 2 )) → Ordered (K 2, List(V 2 )) → List(K 3,V 3 )

10. Causes of Stragglers Internal factorsExternal factors resource capacity of worker nodes is heterogeneous resource competition due to other MapReduce tasks running on the same worker node resource competition due to co-hosted applications input data skew remote input or output source is too slow hardware faulty

11. 0 Outlines 2. Background 3. Previous work 4. Pitfalls 5. Our Design 1. Introduction 6. Evaluation 7. Conclusion

12. Previous work  Google and Dryad  When a stage is close to completion  Backup an arbitrary set of the remaining tasks  Hadoop Original  Backup task whose progress falls behind the average by a fixed gap  LATE (OSDI ’ 08)  Backup task: 1) longest remaining time, 2) progress rate below threshold  Identify worker with its performance score below threshold as slow  Mantri (OSDI’10)  Saving cluster computing resource  Backup up outliers when they show up  Kill-restart when cluster is busy, lazy duplicate when cluster is idle

13. 0 Outlines 2. Background 3. Previous work 4. Pitfalls 5. Our Design 1. Introduction 6. Evaluation 7. Conclusion

14. Pitfalls in Selecting Slow Tasks  Using average progress rate to identify slow tasks and estimate task remaining time  Hadoop and LATE assumes that:  Tasks of the same type process almost the same amount of input data  Progress rate must be either stable or accelerated during a task ’ s lifetime  There are some scenarios that the assumptions will break down

15. Input data skew Sort benchmark on 10GB input data following the Zipf distribution ( =1.0)

16. Phase Percentage varies Different jobs have different phase duration ratio Job in different environments has different phase duration ratio Speed is varying across different phases

17. Reduce Tasks Start Asynchronously Tasks in different phases can not be compared directly

18. Take a Long Time to Identify Straggler Cannot identify straggler in time

19. Pitfalls in Selecting Backup Node  Identifying Slow Worker Node  LATE: Sum of progress of all the completed and running tasks on the node  Hadoop: Average progress rate of all the completed tasks on the node  Some worker nodes may do more time-consuming tasks and get lower performance score unfairly  e.g. doing more tasks with larger amount of data to process or non-local map tasks  Choosing Backup Worker Node  LATE and Hadoop: Ignore data locality  Our observation: a data-local map task can be over three times faster than that of a non-local map task

20. 0 Outlines 2. Background 3. Previous work 4. Pitfalls 5. Our Design 1. Introduction 6. Evaluation 7. Conclusion

21. Selecting Backup Candidates  Using Per-Phase Process Speed  Dividing each task into multiple phases  Using phase process speed to identify slow tasks and estimate task remaining time MapCombineCopySortReduce Map Task:Reduce Task: Map Combine Map Task compare

22. Selecting Backup Candidates  Using EWMA to Predict Process Speed

23. Selecting Backup Candidates  Estimating Task Remaining Time and Backup Time  use the phase average process speed to estimate the remaining time of a phase  To avoid process speed to be fast at the beginning and drop later in copy phase, we estimate the remaining time of copy phase as follows:

24. Selecting Backup Candidates  Maximizing Cost Performance  Cost: the computing resources occupied by tasks  Performance: the shortening of job execution time and the increase of the cluster throughput  We hope that:  when a cluster is idle, the cost for speculative execution is less a concern  When the cluster is busy, the cost is an important consideration

25. Selecting Proper Backup Nodes  Assign backup tasks to the fast nodes  How to measure the performance of nodes?  use predicted process bandwidth of data-local map tasks completed on the node to represent its performance  Consider the data-locality  Note: the process speed of data-local map tasks can be 3 times that of non-local map tasks  Therefore, we keep the process speed statistics of data-local, rack- local, and non-local map tasks for each node  For nodes that do not process any map task on a specific locality level, we use the average process speed of all nodes on this level as an estimate  Launch backup on node i ? If remain time > backup time on node i

26. Summary  A task will be backed up when it meets the following conditions:  it has executed for a certain amount of time (i.e., the speculative lag)  both the progress rate and the process bandwidth in the current phase of the task are sufficiently low  the profit of doing the backup outweighs that of not doing it  its estimated remaining time is longer than the predicted time to finish on a backup node  it has the longest remaining time among all the tasks satisfying the conditions above

27. 0 Outlines 2. Background 3. Previous work 4. Pitfalls 5. Our Design 1. Introduction 6. Evaluation 7. Conclusion

28. Experiment Environment  Two scale:  Small: 30 virtual machines on 15 physical machines  Large: 100 virtual machines on 30 physical machines  Each physical machine:  dual-Processors (2.4GHz Intel(R) Xeon(R) E5620 processor with 16 logic core), 24GB of RAM and two 150GB disks  Organized in three racks connected by 1Gbps Ethernet  Each virtual machine:  2 virtual core, 4GB RAM and 40GB of disk space  Benchmark:  Sort, Wordcount, Grep, Girdmix

29. Scheduling in Heterogeneous Environments  Load of each host in heterogeneous environments LoadHostsVMs 1VMs/host33 2VMs/host1122 5VMs/host15 Total1530

30. Scheduling in Heterogeneous Environments  Working With Different Workloads WorkloadsJob completion time Improvement Cluster throughput Improvement WordCount10%5% Sort19%15% Grep39%38% Gridmix13%15%

31. Scheduling in Heterogeneous Environments  Analysis (using Word Count and Grep) Strategy PrecisionRecall Average find time MapReduc e MapReduceMapReduce Hadoop- LATE 37.6%3%100% 70s66s Hadoop- MCP 45.2%93.3%87.1%100%56s32s Improvement +Accurate prediction +Cost performance All Execution Speed 27%31%39% Cluster Throughput 29%32%38%

32. Scheduling in Heterogeneous Environments  Handling Data Skew (Sort) Execution Speed +37% Cluster Throughput +44% Execution Speed +17% Cluster Throughput +19%

33.  Competing with other applications  Run some I/O intensive processes on some servers  dd process which creates large files in a loop to write random data on some physical machines  MCP can run 36% faster than Hadoop-LATE and increase the cluster throughput by 34%. Scheduling in Heterogeneous Environments

34. Large scale Experiment  Load distribution  MCP finishes jobs 21% faster than Hadoop-LATE and improves the cluster throughput by 16% LoadHostsVMs 3VMs/host2781 5VMs/host420 Total31101

35. Scheduling in Homogeneous Environments  Small scale cluster with each host running 2 VMs  There is no straggler node in the cluster  MCP finishes jobs 6% faster than Hadoop-LATE and 2% faster than Hadoop-None.  Hadoop-LATE behaves worse than Hadoop-None due to too many unnecessary reduce backups  MCP improves reduce backup precision by 40%  MCP can achieve better data locality for map tasks

36. Scheduling Cost  We measure the average time that MCP and Hadoop-LATE spend on speculative scheduling in a job with 350 map tasks and 110 reduce tasks  MCP spends about 0.54ms O(n)  LATE spends 0.74ms O(nlogn)

37. 0 Outlines 2. Background 3. Previous work 4. Pitfalls 5. Our Design 1. Introduction 6. Evaluation 7. Conclusion

38. Conclusion  We provide an analysis of the pitfalls of current speculative execution strategies in MapReduce  Scenarios: data skew, tasks that start asynchronously, improper configuration of phase percentage etc.  We develop a new strategy MCP to handle these scenarios:  Accurate slow task prediction and remaining time estimation  Take the cost performance of computing resources into account  Take both data locality and data skew into consideration when choosing proper worker nodes  MCP fits well in both heterogeneous and homogeneous environments  handle data skew case well, quite scalable, and less overhead

39. Thank You!