1. IMPROVING DATACENTER PERFORMANCE AND ROBUSTNESS WITH MULTIPATH TCP Costin Raiciu†, Sebastien Barre‡, Christopher Pluntke†, Adam Greenhalgh†, Damon Wischik†, Mark Handley†

2. Before Start  我想要先用我的話說一次這篇 Paper 想做的事情 :  這邊 Paper 的重點在於， single path 用在 TCP 上面 的時代已經過去了， multipath TCP 才是主流。  Multipath TCP 又主要有三種 topologies: fattree, VL2, Bcube 。  這篇 paper 又分析了在不同情況下 multipath 確實 比 single path 優秀的地方。還有一些影響 performance 的因素的分析。  最後是有沒有其他可能再發展的 topology

3. OUTLINE  Introduction  Data center networking  Topology.  Routing.  Path Selection.  Congestion Control.  Multipath TCP in summary  MPTCP in data center  Examples of Benefits  Analysis Influence of Topology Number of Subflows Influence of the Traffic Matrix  Evolving topologies with MPTCP  Analysis  Discussion about DHFT  Summary

4. Introduction  Traditional single path TCP has many significant problems, ill-suited for such network. (intra-data center traffic.)  Ex. Bottleneck, can’t scale up, congestion.  We propose using Multipath TCP as a replacement for TCP in such data centers.

5. Introduction  Traditional data center topology. …

6. Introduction  4 components to a data center architecture:  1. physical topology  2. routing  3. path selection  4. congestion control of traffic on the selected path.

7. OUTLINE  Introduction  Data center networking  Topology.  Routing.  Path Selection.  Congestion Control.  Multipath TCP in summary  MPTCP in data center  Examples of Benefits  Analysis Influence of Topology Number of Subflows Influence of the Traffic Matrix  Evolving topologies with MPTCP  Analysis  Discussion about DHFT  Summary

8. Data center networking---Topology  Traditional data center topology. …

9. Data center networking---Topology  That has many problems and can be improved a lot like we have discussed before.  So, fat-tree is one kind of solutions.

10. Fat Tree Topology [Fares et al., 2008; Clos, 1953] K=4 Aggregation Switches K Pods with K Switches each Racks of servers

11. Data center networking---Topology  And we also have VL2 and Bcube these two kinds of solutions.

12. BCube Topology [Guo et al, 2009] BCube (4,1)

13. Data center networking---Topology

14. VL2 Topology [Greenberg et al, 2009, Clos topology] 10Gbps 20 hosts 10Gbps …

15. Data center networking---Routing  We cannot expect the host itself to know which of these paths is the least loaded, so the routing system must spread traffic across these paths.  Use randomized load balancing, where each flow is assigned a random path from the set of possible paths.

16. Data center networking--- Path Selection  randomized load balancing 有他一定的問題，就 是有一定的機率某些 path 會有超大流量、同時 會有一些 path 沒什麼 ( 極低或零 ) 流量。  The use of a centralized flow scheduler has been proposed.

18. Multipath TCP: Congestion Control [NSDI, 2011]

19. Data center networking--- Congestion Control  Our hypothesis is that given sufficiently many randomly chosen paths, MPTCP will find at least one good unloaded path, and move most of its traffic that way. In so doing it will relieve congestion on links that got more than their fair share of ECMP balanced flows.

20. Multipath TCP in summary  By making the window increase depend on the total window size, subflows that have large windows increase faster than subflows with small windows.

21. OUTLINE  Introduction  Data center networking  Topology.  Routing.  Path Selection.  Congestion Control.  Multipath TCP in summary  MPTCP in data center  Examples of Benefits  Analysis Influence of Topology Number of Subflows Influence of the Traffic Matrix  Evolving topologies with MPTCP  Analysis  Discussion about DHFT  Summary

22. MPTCP in data center  Three main benefits:  1. Better aggregate throughput  2. Better fairness  3. Better robustness

23. MPTCP in data center  Depend on 4 requirements:  1. The congestion control scheme used.  2. The physical topology.  3. The traffic matrix generated by the applications.  4. The level of load in the network.

24. Examples of Benefits  Throughput:  Using MPTCP in a FatTree network, we can see the more subflows we have, the more throughput we have.  The reason is simple, n flows share one link, each flow need to achieves 1/n capacity.

25. Examples of Benefits

26.  Fairness:  Every host’s throughput is shown ranked in order of increasing throughput. Is is clear that not only did the utilization improve with MPTCP, but also the fairness improved.

27. Examples of Benefits  這是一個累積百分比的圖，可以看出 single TCP 有些 path perform 真的很差 ( 很閒的意思 ) ，而 MPTCP 則大幅改善此情況。

28. Analysis  Then, we have questions:  How many subflows are needed?  How does the topology affect results?  How does the traffic matrix affect results?

29. Analysis  We need to view in these aspects to answer the question:  Influence of Topology  Number of Subflows  Influence of the Traffic Matrix

30. Influence of Topology  We compare VL2 and Bcube.

31. Influence of Topology  Why Bcube is relative low ？  I guess, just guess, is because in Bcube, pathes between two hosts are not in same length which lead to worse flows management.

32. Number of Subflows  We were surprised that eight subflows were needed to achieve 90% throughput.  Why eight, and on what does this depend?

33. Number of Subflows  Results:

34. Influence of the Traffic Matrix

35. OUTLINE  Introduction  Data center networking  Topology.  Routing.  Path Selection.  Congestion Control.  Multipath TCP in summary  MPTCP in data center  Examples of Benefits  Analysis Influence of Topology Number of Subflows Influence of the Traffic Matrix  Evolving topologies with MPTCP  Analysis  Discussion about DHFT  Summary

36. Evolving topologies with MPTCP  In single homed topologies:  Hosts links are often bottlenecked.  ToR switches failure also lead to crash tens of hosts.

37. Evolving topologies with MPTCP  以 FatTree 為例子

38. Evolving topologies with MPTCP  So, how do we improved ？  There are two solutions: perfect switch and dual- homed FatTree.

39. Evolving topologies with MPTCP  Perfect switch:  Serves as a good control experiment, giving an upper bound on what any network core might provide using single links to the hosts.

40. Evolving topologies with MPTCP  Dual-homed FatTree (DHFT):  is to the host and four connect the links between the two layers of switches.  If we remove one port per host from the core and use it to connect the second interface on each server, the network requires the same number of switch ports.

41. Analysis  QUESTION:  Will perfect switch really improve performance ？

42. Analysis  Testing throughput over MPTCP, perfect switch TCP, normal TCP.

43. Analysis  QUESTION:  Will perfect switch really improve performance ？  ANSWER:  Yes, it is.

44. Analysis  QUESTION:  Dual-homed FatTree really work ？  Does it have any problem ？ When may troubles occur ？

45. Analysis  Testing relative throughput over MPTCP, TCP DHFT, TCP perfect switch.

46. Analysis  QUESTION:  Dual-homed FatTree really work ？  Does it have any problem ？ When may troubles occur ？  ANSWER:  Yes, it works only when load is low. When load is high, throughput will go down near to TCP.

47. Discussion about DHFT  DHFT’s worst case performance is 75% and best case is around 200%.  在流量不穩定的情況下， DHFT + MPTCP 有可能會有 較差的表現產生。  Beyond performance, DHFT improves robustness: any lowerpod switch failure does not cut-off an entire rack of servers.  DHFT is not optimal by any measure, but it shows that we can create topologies with better performance if we assume MPTCP is the transport protocol.

48. OUTLINE  Introduction  Data center networking  Topology.  Routing.  Path Selection.  Congestion Control.  Multipath TCP in summary  MPTCP in data center  Examples of Benefits  Analysis Influence of Topology Number of Subflows Influence of the Traffic Matrix  Evolving topologies with MPTCP  Analysis  Discussion about DHFT  Summary

49. Summary  “One flow, one path” thinking has constrained datacenter design  Collisions, unfairness, limited utilization  Multipath transport enables resource pooling in datacenter networks:  Improves throughput  Improves fairness  Improves robustness

50. Reference  conferences.sigcomm.org/sigcomm/2011/slides/s2 66.ppt ‎

51.  Q & (A)