1. A Case for Performance- Centric Network Allocation Gautam Kumar, Mosharaf Chowdhury, Sylvia Ratnasamy, Ion Stoica UC Berkeley

2. Datacenter Applications

3. 3 Data Parallelism Applications execute in several computation stages and require transfer of data between these stages (communication). Computation in a stage is split across multiple nodes. Network has an important role to play, 33% of the running time in Facebook traces. (Orchestra, SIGCOMM 2011) M M M M R R R R J J M M R R M M R R J J M M R R J J Map Reduce Join (*RoPE, NSDI 2012) * HotCloud June 12, 2012

4. 4 Data Parallelism Users, often, do not know what network support they require.  Final execution graph created by the framework. Frameworks know more, provide certain communication primitives.  e.g., Shuffle, Broadcast etc. HotCloud June 12, 2012

5. 5 Scope Private clusters running data parallel applications. Little concern for adversarial behavior. Application level inefficiencies dealt extrinsically. HotCloud June 12, 2012

6. Current Proposals

7. 7 Explicit Accounting Virtual cluster based network reservations. (Oktopus, SIGCOMM 2011) Time-varying network reservations. (SIGCOMM 2012) DRAWBACK: Exact network requirements often not known; non work- conserving. HotCloud June 12, 2012

8. 8 Fairness-Centric Flow level fairness or Per-Flow. (TCP) Fairness with respect to the sources. (Seawall, NSDI 2012) Proportionality in terms of total number of VMs. (FairCloud, SIGCOMM 2012) DRAWBACK: Gives little guidance to developers about the performance they can expect while scaling their applications. HotCloud June 12, 2012

9. 9 In this work... A new perspective to share the network amongst data-parallel applications – performance-centric allocations:  enabling users to reason about the performance of their applications when they scale them up.  enabling applications to effectively parallelize to preserve the intuitive mapping between scale-up and speed-up. Contrast / relate performance-centric proposals with fairness-centric proposals. HotCloud June 12, 2012

10. Performance- Centric Allocations

11. 1 λ λ/2 λ λ λ Shuffle Broadcast Types of Transfers* (*Orchestra, SIGCOMM 2011) HotCloud June 12, 2012

12. 1212 λ λ λ λ λ λ/2 λ λ λ 2λ λ/2 2X Scale UP Total Data = 4λ Total Data = 2λ λ/2 ShuffleBroadcast Scaling up the application HotCloud June 12, 2012

13. 1313 Performance -Centric Allocations Understand the support that the application needs from the network to effectively parallelize. At a sweet spot – framework knows application’s network requirements. HotCloud June 12, 2012

14. 1414 Shuffle-only clusters λ/2 λ 2λ AmAm ArAr BmBm BrBr λ BmBm BrBr λ/2 tAmtAm tAstAs tArtAr tBmtBm tBrtBr tBstBs HotCloud June 12, 2012

15. 1515 Shuffle-only Clusters λ/2 λ 2λ AmAm ArAr BmBm BrBr λ BmBm BrBr λ/2 tAmtAm t A s = 2λ/α tArtAr t A m /2t A r /2t B s = λ/2α = t A s /4 α α t B < t A /2 λ/2 λ 2λ AmAm ArAr BmBm BrBr λ BmBm BrBr λ/2 tAmtAm t A s = 2λ/α tArtAr t A m /2 t A r /2t B s = λ/α = t A s /2 α α/2 t B = t A /2 Per-FlowProportional HotCloud June 12, 2012

16. 1616 Broadcast-only Clusters λ λ 2λ AmAm ArAr BmBm BrBr λ BmBm BrBr λ λ λ tAmtAm tAstAs tArtAr tBmtBm tBrtBr tBstBs HotCloud June 12, 2012

17. 1717 Broadcast-only Clusters λ λ 2λ AmAm ArAr BmBm BrBr λ BmBm BrBr λ λ λ tAmtAm t A s = 2λ/α tArtAr t A m /2t A r /2t B s = 2λ/α = t A s α α/2 t B > t A /2 λ 2λ AmAm ArAr BmBm BrBr λ BmBm BrBr tAmtAm t A s = 2λ/α tArtAr t A m /2t A r /2t B s = λ/α = t A s /2 t B = t A /2 α α λ λ λ λ Proportional Per-Flow HotCloud June 12, 2012

18. 1818 Recap TCP in shuffle gives more than requisite speed-up and thus hurts performance of small jobs. Proportionality achieves the right balance. Proportionality in broadcast limits parallelism. TCP achieves the right balance. Degree of Parallelism Speed Up TCP (Shuffle) Prop. (Shuffle) TCP. (Broadcast) Prop. (Broadcast) HotCloud June 12, 2012

19. 1919 Complexity of a transfer x N -transfer if x is the factor by which the amount of data transferred increases when a scale up of N is done, x  [1, N]. Shuffle is a 1 N -transfer and broadcast is an N N -transfer. Performance-centric allocations encompass x. HotCloud June 12, 2012

20. Heterogeneous Frameworks and Congested Resources

21. 2121 Share given based on the complexity of the transfer. The job completion time of both jobs degrades uniformly in the event of contention. Both finish in 6s Both finish in 4s 2G AmAm ArAr BmBm BrBr 1G A’ m A’ r 1G A’ m A’ r 0.5G 1G1G B’ m B’ r 1G1G B’ m B’ r 500Mbps 2Gb to send 500Mbps 2Gb to send ~333Mbps 2Gb to send ~666Mbps 4Gb to send 2X Scale UP 0.5G 1G HotCloud June 12, 2012

22. 2 Network Parallelism Isolation between the speed-up due to the scale- up for the application and the performance degradation due to finite resources. y’ y N (α)(α) X α : degradation due to limited resources y : old running time y’: new running time after a scale-up of N HotCloud June 12, 2012

23. Summary

24. 2424 Understand performance-centric allocations and their relationship with fairness-centric proposals.  Proportionality is the performance-centric approach for shuffle-only clusters.  Breaks down for broadcasts, per-flow is the performance-centric approach for broadcast-only clusters. An attempt to a performance-centric proposal for heterogeneous transfers.  Understand what happens when resources get congested. HotCloud June 12, 2012

25. Future Work

26. 2626 A more rigorous formulation.  Some questions to be answered: different N 1 and N 2 on both sides of the stage etc. Analytical and experimental evaluation of the policies.  Whether redistribution of completion time or total savings. HotCloud June 12, 2012

27. Thank you