Computer Networks COSC 6377

Transcription

1 Computer Networks COSC 6377 Lecture 25 Fall 2011 November 30,

2 Announcements Grades will be sent to each student for verificagon P2 deadline extended 2

3 Large- scale computagon Search Engine Tasks Crawl Build Index Answer query How much computagon, storage, and bandwidth do we need? How to get them cheap? Economy of scale 3

4 Map Reduce Large list of tasks that can be performed in parallel Count the number of web pages in which a special word appears Count the number of Gmes a person appears in a collecgon of images Map Reduce Transform each element (map) Aggregate (reduce) 4

5 Internet- scale Map Reduce Large number of workers performing map and reduce operagon CommunicaGon between machines Results of map needs to be processed by reduce In general, output of one operagon in a machine an input for another process in a different machine How to opgmize data center topology for communicagon between arbitrary machines? Trees, and other topologies, we will see later today 5

6 Data center video 6

7 Rest of the slides are from CosGn Raiciu s SIGCOMM talk 7

8 Improving Datacenter Performance and Robustness with Multipath TCP Costin Raiciu Department of Computer Science University Politehnica of Bucharest Sebastien Barre (UCL-BE), Christopher Pluntke (UCL), Adam Greenhalgh (UCL), Damon Wischik (UCL) and Mark Handley (UCL) Thanks to:

9 Motivation Datacenter apps are distributed across thousands of machines Want any machine to play any role To achieve this: Use dense parallel datacenter topologies Map each flow to a path Problem: This is the wrong place to start Naïve random allocation gives poor performance Improving performance adds complexity

10 Contributions Multipath topologies need multipath transport Multipath transport enables better topologies

11 To satisfy demand, modern datacenters provide many parallel paths Traditional Topologies are treebased Poor performance Not fault tolerant Shift towards multipath topologies: FatTree, BCube, VL2, Cisco, EC2

12 Fat Tree Topology [Fares et al., 2008; Clos, 1953] K=4 AggregaGon Switches 1Gbps 1Gbps K Pods with K Switches each Racks of servers

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

14 Collisions

15 Single-path TCP collisions reduce throughput 1000 Throughput (Mb/s) Optimal Throughput TCP Flow Throughput Rank of Flow

16 Collision

17

18

19 Not fair

20 Not fair

21

22

23 No matter how you do it, mapping each flow to a path is the wrong goal

24 Instead, we should pool capacity from different links

25 Instead, we should pool capacity from different links

26 Instead, we should pool capacity from different links

27 Instead, we should pool capacity from different links

28 Multipath Transport

29 Multipath Transport can pool datacenter networks Instead of using one path for each flow, use many random paths Don t worry about collisions. Just don t send (much) traffic on colliding paths

30 Multipath TCP Primer [IETF MPTCP WG] MPTCP is a drop in replacement for TCP MPTCP spreads application data over multiple subflows

31 Multipath TCP: Congestion Control [NSDI, 2011]

32 MPTCP better utilizes the FatTree network 1000 Throughput (Mb/s) MPTCP Optimal Throughput TCP Flow Throughput Rank of Flow

33 MPTCP on EC2 Amazon EC2: infrastructure as a service We can borrow virtual machines by the hour These run in Amazon data centers worldwide We can boot our own kernel A few availability zones have multipath topologies 2-8 paths available between hosts not on the same machine or in the same rack Available via ECMP

34 Amazon EC2 Experiment 40 medium CPU instances running MPTCP For 12 hours, we sequentially ran all-to-all iperf cycling through: TCP MPTCP (2 and 4 subflows)

35 MPTCP improves performance on EC2 Throughput (Mb/s) TCP MPTCP, 4 subflows MPTCP, 2 subflows Same Rack Flow Rank

36 What do the benefits depend on?

37 How many subflows are needed? How does the topology affect results? How does the traffic matrix affect results?

38 At most 8 subflows are needed Throughput (% of optimal) Total Throughput TCP RLB Multipath TCP

39 MPTCP improves fairness in VL topologies VL2 Throughput (Mb/s) Fairness is important: 400 Jobs finish when the slowest worker finishes Single Path TCP MPTCP, 2 subflows MPTCP, 4 subflows Rank of Flow

40 MPTCP improves throughput and fairness in BCube Throughput (Mb/s) BCube Single Path TCP MPTCP, 2 subflows MPTCP, 5 subflows Rank of Flow

41 Oversubscribed Topologies n To saturate full bisectional bandwidth: There must be no traffic locality All hosts must send at the same time Host links must not be bottlenecks n It makes sense to under-provision the network core This is what happens in practice Does MPTCP still provide benefits?

42 Performance improvements depend on Relative MPTCP Throughput Underloaded traffic matrix TCP MPTCP Sweet Spot Overloaded Connections per host Increase Load

43 What is an optimal datacenter topology for multipath transport?

44 In single homed topologies: o Hosts links are often bottlenecks o ToR switch failures wipe out tens of hosts for days MulG- homing servers is the obvious way forward

45 Fat Tree Topology

46 Fat Tree Topology Upper Pod Switch ToR Switch Servers

47 Dual Homed Fat Tree Topology Upper Pod Switch ToR Switch Servers

48 Is DHFT any better than Fat Tree? Not for traffic matrices that fully utilize the core Let s examine random traffic patterns Other TMs in the paper

49 DHFT provides significant improvements when core is not overloaded Relative throughput TCP Perfect Switch MPTCP DHFT TCP DHFT Core Underloaded Core Overloaded Connections per host

50 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 One flow, many paths frees designers to consider topologies that offer improved performance for similar cost

51 Backup Slides

52 Effect of MPTCP on short flows Flow sizes from VL2 dataset MPTCP enabled for long flows only (timer) Oversubscribed Fat Tree topology Results: TCP/ECMP Completion time: 79ms Core Utilization: 25% MPTCP 97ms 65%

53 MPTCP vs Centralized Dynamic Scheduling Throughput (% of max) Infinite Centralized Scheduling MPTCP VLB 1s 500ms 100ms 10ms MTCP Scheduling First Interval Fit Scheduler

54 Centralized Scheduling: Setting the Threshold Throughput 1Gbps Hope 17% worse than mulgpath TCP 100Mbps App Limited

55 Centralized Scheduling: Setting the Threshold Throughput 1Gbps 21% worse than mulgpath TCP 100Mbps App Limited Hope

56 Centralized Scheduling: Setting the Threshold Throughput 1Gbps 500Mbps 100Mbps 51% 17% 45% 21%

57 Effect of Locality in the Dual Homed Fat Tree Total Network Throughput (Gb/s) Perfect Switch TCP MPTCP Multi-path region Single-path region Locality (percent)

58 Overloaded Fat Tree: better fairness with Multipath TCP

59 VL2 Topology [Greenberg et al, 2009, Clos topology] 10Gbps 10Gbps 20 hosts

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