1. Orchestra Managing Data Transfers in Computer Clusters Mosharaf Chowdhury, Matei Zaharia, Justin Ma, Michael I. Jordan, Ion Stoica UC Berkeley

2. Moving Data is Expensive Typical MapReduce jobs in Facebook spend 33% of job running time in large data transfers Application for training a spam classifier on Twitter data spends 40% time in communication 2

3. Limits Scalability Scalability of Netflix-like recommendation system is bottlenecked by communication 3 Did not scale beyond 60 nodes »Comm. time increased faster than comp. time decreased

4. Transfer Patterns Transfer: set of all flows transporting data between two stages of a job »Acts as a barrier Completion time: Time for the last receiver to finish Map Shuffle Reduce Broadcast Incast* 4

5. Contributions 1.Optimize at the level of transfers instead of individual flows 2.Inter-transfer coordination 5

6. TC (broadcast) HDFS Tree Cornet HDFS Tree Cornet HDFS Tree Cornet HDFS Tree Cornet TC (shuffle) Hadoop shuffle WSS Hadoop shuffle WSS shuffle broadcast 1broadcast 2 ITC Fair sharing FIFO Priority Fair sharing FIFO Priority 6 Shuffle Transfer Controller (TC) Shuffle Transfer Controller (TC) Broadcast Transfer Controller (TC) Broadcast Transfer Controller (TC) Broadcast Transfer Controller (TC) Broadcast Transfer Controller (TC) Inter-Transfer Controller (ITC) Inter-Transfer Controller (ITC) Orchestra

7. Outline Cooperative broadcast (Cornet) »Infer and utilize topology information Weighted Shuffle Scheduling (WSS) »Assign flow rates to optimize shuffle completion time Inter-Transfer Controller »Implement weighted fair sharing between transfers End-to-end performance 7 TC (broadcast) HDFS Tree Cornet HDFS Tree Cornet HDFS Tree Cornet HDFS Tree Cornet TC (shuffle) Hadoop shuffle WSS Hadoop shuffle WSS ITC Fair sharing FIFO Priority Fair sharing FIFO Priority ITC Shuffle TC Shuffle TC Broadcast TC Broadcast TC Broadcast TC Broadcast TC

8. Cornet : Cooperative broadcast ObservationsCornet Design Decisions 1.High-bandwidth, low-latency network Large block size (4-16MB) 2.No selfish or malicious peers No need for incentives (e.g., TFT) No (un)choking Everyone stays till the end 3.Topology matters Topology-aware broadcast 8 Broadcast same data to every receiver »Fast, scalable, adaptive to bandwidth, and resilient Peer-to-peer mechanism optimized for cooperative environments

9. Cornet performance 9 1GB data to 100 receivers on EC2 4.5x to 5x improvement Status quo

10. Topology-aware Cornet Many data center networks employ tree topologies Each rack should receive exactly one copy of broadcast »Minimize cross-rack communication Topology information reduces cross-rack data transfer »Mixture of spherical Gaussians to infer network topology 10

11. Topology-aware Cornet 11 ~2x faster than vanilla Cornet 200MB data to 30 receivers on DETER 3 inferred clusters

12. Status quo in Shuffle 12 r1r1 r1r1 r2r2 r2r2 s2s2 s2s2 s3s3 s3s3 s4s4 s4s4 s1s1 s1s1 s5s5 s5s5 Links to r 1 and r 2 are full: Link from s 3 is full: Completion time: 3 time units 2 time units 5 time units

13. Allocate rates to each flow using weighted fair sharing, where the weight of a flow between a sender-receiver pair is proportional to the total amount of data to be sent 13 Up to 1.5X improvement Completion time: 4 time units Weighted Shuffle Scheduling r1r1 r1r1 r2r2 r2r2 s2s2 s2s2 s3s3 s3s3 s4s4 s4s4 s1s1 s1s1 s5s5 s5s5 112211

14. Inter-Transfer Controller aka Conductor Weighted fair sharing »Each transfer is assigned a weight »Congested links shared proportionally to transfers weights Implementation: Weighted Flow Assignment (WFA) »Each transfer gets a number of TCP connections proportional to its weight »Requires no changes in the network nor in end host OSes 14

15. Benefits of the ITC Two priority classes »FIFO within each class Low priority transfer »2GB per reducer High priority transfers »250MB per reducer Shuffle using 30 nodes on EC2 43% reduction in high priority xfers 6% increase of the low priority xfer Without Inter-transfer Scheduling Priority Scheduling in Conductor

16. End-to-end evaluation Developed in the context of Spark – an iterative, in- memory MapReduce-like framework Evaluated using two iterative applications developed by ML researchers at UC Berkeley »Training spam classifier on Twitter data »Recommendation system for the Netflix challenge 16

17. Faster spam classification 17 Communication reduced from 42% to 28% of the iteration time Overall 22% reduction in iteration time

18. Scalable recommendation system Before After 18 1.9x faster at 90 nodes

19. Related work DCN architectures (VL2, Fat-tree etc.) »Mechanism for faster network, not policy for better sharing Schedulers for data-intensive applications (Hadoop scheduler, Quincy, Mesos etc.) »Schedules CPU, memory, and disk across the cluster Hedera »Transfer-unaware flow scheduling Seawall »Performance isolation among cloud tenants 19

20. Summary Optimize transfers instead of individual flows »Utilize knowledge about application semantics Coordinate transfers »Orchestra enables policy-based transfer management »Cornet performs up to 4.5x better than the status quo »WSS can outperform default solutions by 1.5x No changes in the network nor in end host OSes 20 http://www.mosharaf.com/

21. BACKUP SLIDES 21

22. MapReduce logs Weeklong trace of 188,000 MapReduce jobs from a 3000-node cluster Maximum number of concurrent transfers is several hundreds 22 33% time in shuffle on average

23. Monarch ( Oakland11 ) Real-time spam classification from 345,000 tweets with urls »Logistic Regression »Written in Spark Spends 42% of the iteration time in transfers »30% broadcast »12% shuffle 100 iterations to converge 23

24. Collaborative Filtering Netflix challenge »Predict users ratings for movies they havent seen based on their ratings for other movies 385MB data broadcasted in each iteration 24 Does not scale beyond 60 nodes

25. Cornet performance 25 1GB data to 100 receivers on EC2 4.5x to 6.5x improvement

26. Shuffle bottlenecks An optimal shuffle schedule must keep at least one link fully utilized throughout the transfer 26 At a senderAt a receiverIn the network

27. Current implementations 27 Shuffle 1GB to 30 reducers on EC2