Tarazu Optimizing MapReduce On Heterogeneous Clusters

1. TarazuOptimizing MapReduce On Heterogeneous Clusters 72130310 임규찬

2. 목차 • Abstract of Paper • Abstract of paper • Reference of paper – LATE • Introduction • Issue with Heterogeneity • Tarazu • Experimental Result

3. Abstract of Paper • Heterogeneous Cluster환경에서 MapReduce기법의 최적화를 연구함. • 데이터 센터 규모의 클러스터 환경에서 경제적 이유로Heterogeneous를 도입하고 있음. • MapReduce기법을 통한 BigData처리가 가능해짐. • 기존의 기법으로는 성능이 오히려 떨어졌음. • Straggler task Managing이용한 기존 연구는 효과 없음 • 그 예시로써 Improving MapReduce Performance in Heterogeneous Environments 논문을 비교함.

4. Reference of PaperImproving MapReduce Performance in Heterogeneous Environments • Straggler Task제어를 통한 Heterogeneous 최적화 • Node is available but is performing poorly Condition • Can arise many reason, faulty hardware and misconfiguration • LATE Scheduler 제안 • Longest Approximate Time to End • Task별 Progress rate를 이용함 • P rogressScore/Amount of time the task • Unfortunately, LATE alone is not sufficient to address hardware heterogeneity.

5. Introduction - Tarazu(तराजू) • ‘균형’을 뜻하는 힌디어 • MapReduce연산에 있어서 균형을 추구하도록 설계

6. Introduction -MapReduce • 대용량 데이터를 분산 컴퓨팅 환경에서 병렬처리 하도록 만들어진 프레임워크 • Homogeneous cluster에 최적화.

7. Introduction -Heterogeneous Computing • 서로 다른 코어로 이루어진 시스템을 이용한 Computing • CPU/GPU를 이용한 GPGPU • CPU/GPU 각각의 장점을 극대화하여 성능 향상을 꾀함. • OpenCL, CUDA, DirectCompute등 존재. • 본 논문에서 다루지 않음 • High/Row Node를 이용한 Clustering • 전력, 가격 등 금전적인 요소에서의 최적화 • 본 논문에서 10개의 Xeon Node, 80개의 Atom Node사용

8. Issue with Heterogeneity-Background : MapReduce • Four phase Excution Model • Map computation • produces <Key, Value> tuple • Shuffle • all Map to all Reduce personalized Communication • Sorts • Grouping all the tuples for same Key • Reduce computation • Processes all the tuples for a key & produce final output

9. Issue with Heterogeneity-Reasons for poor performance on heterogeneous clusters • Dynamic Load-balancing in MapReduce • Slower nodes fewer tasks/faster nodes more tasks • Heterogeneity is slow than Homogeneity • 20-75% slower for six out of eleven benchmarks. • Heterogeneity can be degrades performance • Poor performance is due to two Key factors • Non-intuitive • Other intuitive

10. Issue with Heterogeneity-Reasons for poor performance on heterogeneous clusters • Factor 1 : Non-intuitive • Interaction between load balancing and network traffic • In Heterogeneous, cause remote task • Xeon is fast, Atom is slow. So Xeon stole Atom task • Remote task can Network Traffic • Network Traffic is exacerbated heavy Shuffle

11. Issue with Heterogeneity-Reasons for poor perfermance on heterogeneous clusters • Factor 2 : Intuitive • Reduce phase imbalance amplified by heterogeneity • Reduce phase load imbalance • Different processing speeds cause long time

12. Issue with Heterogeneity-A Simple(?) analytical model Map Finish Time (High/Low System중 Map 연산이 늦게 끝나는 시간값) Number of input data in bisection (Remote Task로 인한 데이터 + 셔플 데이터) Shuffle Finish Time (Remote task로 인한 시간 혹은 MFT) Reduce Finish Time (Remote task로 인한 시간 혹은 MFT)

13. Tarazu • Two problems in MapReduce • Map-side built-in load balancing results in remote Map • Reduce-side load imbalance across the nodes • Tarazu consist of three components • Communication-Aware Load Balancing of Map computation • Communication-Aware Scheduling of Map computation • Predictive Load Balancing of Reduce computation

14. Tarazu- Communication-Aware Load Balancing of Map computation • Based on key observation • Due to the overlap between Map computation and Shuffle • In Shuffle is critical, ‘no-steal mode’ • Pick up remote task when Shuffle end • There are no remote Map tasks to compete with Shuffle • Reduce the I/O Processing overhead • Slower nodes perform more work

15. Tarazu- Communication-Aware Load Balancing of Map computation • In Map Computation is Critical, ‘task-steal mode’ • Concern of CAS. • CALB’s mode change using shuffleLag • Using MapReduce monitorfor fault tolerance • Diffence of number of • Map task that have completed their computation • Have completed their communication in all nodes • Deciding the Source of criticality once is enough without repeated, dynamic check.

16. Tarazu- Communication-Aware Scheduling of Map computation • Determine how many remote tasks needed • Using in CALB ‘task-steal’ mode • Using to avoid increase SFT • To avoid traffic, CAS spreads out the remote task by interleaving them with local task

17. Tarazu- Communication-Aware Scheduling of Map computation • CAS has other benefits • By interleaving remote tasks with local tasks, CAS achieves better overlap between remote task communication and local task computation on both sender and receiver sides • Remote tasks read input data faster by avoiding bursts

18. Tarazu- Predictive Load Balancing of Reduce computation • Better load balance in the Reduce phase • Skewing the intermediate key distribution • Reduce max term RFT • Each Reduce task save number of fast/slow nodes.

19. Experimental Methodology • Using Heterogeneous Cluster Environment • 10 Xeon-based/80 Atom-based server nodes • Using Hadoop 0.20.2 • Compare another solution, LATE

20. Experimental Result-Performance • Heterogeneous 기법을 통한 시스템 장점 극대화 • Shuffle-Critical의 경우에는 Atom의 물량 반영 • Map-Critical의 경우에는 Xeon의 성능 반영

21. Experimental Result-Effect of CALB, CAS and PLB

22. Experimental Result-Sensitivity to extent of heterogeneity

23. Experimental Result-Effect of skewed input data dist.

24. Reference • Improving MapReduce Performance in Heterogeneous Environments –University of California, Berkeley • https://developers.google.com/appengine/docs/python/dataprocessing/ • http://www.cpubenchmark.net/