1. 6.888 Lecture 6: Network Performance Isolation Mohammad Alizadeh Spring 2016 1

2. Multi-tenant Cloud Data Centers Shared infrastructure between multiple tenants/apps

3. Lack of Performance Predictability Unpredictable performance, esp. at the tail GAE memcache read 100 values 3

4. Congestion Kills Predictability 4 Apr 20134NSDI 2013

5. 5 ?

6. Sharing the Network 6 Alice’s Switch VM1 VM2 VMn VM3 … Bob’s Switch VM1 VM2 VMi VM3 … Customer specifies capacity of the virtual NIC. No traffic matrix. … … Hose Model (Duffield et al., SIGCOMM’99) 2Ghz VCPU 15GB memory 1Gb/s network

7. Sharing the Network Tenant selects bandwidth guarantees. Models: Hose, VOC, TAG Place VMs, ensuring all guarantees can be met Enforce bandwidth guarantees & Provide work-conservation VM setup Runtime Oktopus [SIGCOMM’10] Hadrian [NSDI’13] CloudMirror [SIGCOMM’14] Seawall [NSDI’10] FairCloud [SIGCOMM’12] EyeQ [NSDI’13] ElasticSwitch [SIGCOMM’13] ….  Adapted from slide by Lucian Popa

8. Example Runtime System: EyeQ (NSDI’13) 8

9. Shim Distributed Rate Allocation VM 2Gb/s 8Gb/s 2Gb/s 8Gb/s Shim 10Gb/s pipe (min) Rate Guarantees EyeQ Shim Layer In the trusted Domain (Hypervisor/NIC)

10. Distributed Rate Allocation VM 2Gb/s 8Gb/s 2Gb/s 8Gb/s 5Gb/s 10Gb/s pipe (min) Rate Guarantees

11. RX Module Distributed Rate Allocation VM 2Gb/s 8Gb/s 2Gb/s 8Gb/s 5Gb/s

12. Distributed Rate Allocation VM 2Gb/s 8Gb/s 2Gb/s 8Gb/s 1Gb/s 8Gb/s

13. VM 2Gb/s 8Gb/s 2Gb/s 8Gb/s 1Gb/s 8Gb/s 5Gb/s Distributed Rate Allocation

14. RX Module Distributed Rate Allocation VM 2Gb/s 8Gb/s 2Gb/s 8Gb/s 1Gb/s 5Gb/s Spare capacity

15. Distributed Rate Allocation VM 2Gb/s 8Gb/s 2Gb/s 8Gb/s 2.5Gb/s 5Gb/s

16. Transmit/Receive Modules VM 2Gb/s 8Gb/s 1Gb/s Congestion detectors Rate limit. RCP: Rate feedback (R) every 10kB (no per-source state needed) Per-destination rate limiters: only if dest. is congested… bypass otherwise Feedback pkt Rate: 1Gb/s 2Gb/s 8Gb/s

17. Sharing the Network Tenant selects bandwidth guarantees. Models: Hose, VOC, TAG Place VMs, ensuring all guarantees can be met Enforce bandwidth guarantees & Provide work-conservation VM setup Runtime  Adapted from slide by Lucian Popa Cloud Mirror Uses ElasticSwitch [SIGCOMM’13]

18. Cloud Mirror 18  Slides based on presentation by JK Lee (HP)

19. Motivation Cloud applications are diverse & complex Bandwidth models like pipe and hose not a good fit 19 [Bing.com traffic pattern, Sigcomm’12] web DB cache web logic

20. Hose model is unfit Hose aggregates BW towards different components – Too coarse-grained – Prevents accurate and efficient guarantees on infrastructure intra-component (self-edge) inter-component

21. Hose is too coarse-grained web logicDB Web … … Logic DB 400 100 300 200 500 800 TCP-like fair allocation would yield 300:200 3-tier web example Hose model congestion

22. ww 2B web (N) B … LL … DD … logic (N) DB (N) 2B Hose over-provisions physical link bandwidth Hose model reservation at L 2 : 2B · N N: # VMs in each tier B: per-VM per-edge bandwidth Physical deployment example 2X overprovision by Hose Model 2 B N logic - DB demand = B · N web (N) logic (N) DB (N) BB B web + logic DB L1L1 L2L2 ww … LL … DD …

23. Contributions 1.Tenant Application Graph (TAG) - Accurate for complex apps - Flexible to elastic scaling - Intuitive 2.VM Placement Algorithm - Guarantee bandwidth and high availability - Efficient for network and compute resources 23

24. Tenant Application Graph (TAG) 1. Aggregate pipes (like Hose) - Model simplicity - Multiplexing gain 2. Preserve inter-component structure (like Pipe) - Accurately capture application demands - Efficiently utilize network resources DB mem web logic DB mem web logic DB mem web logic Component-level graph

25. Tenant Application Graph (TAG) web (N w ) DB (N D ) B snd B rcv B in web DB B snd B rcv TAG model B snd = per-VM sending bandwidth (VM-to-component aggregation) B rcv = per-VM receiving bandwidth (component-to-VM aggregation) What do self-edges mean?

26. Abstract models in TAG Self-edge  Hose Directional edge  directional Hose, Virtual Trunk Total guarantee of virtual trunk = min(B snd ·N w, B rcv ·N D ) B rcv web(N w ) B sn d … … DB(N D ) B in Virtual Switch Virtual Trunk web (N w ) DB (N D ) B snd B rcv B in TAG model

27. Questions How are TAGs constructed? How to predict bandwidth demands? What is missing for the TAG model?

28. CloudMirror operation VM placement BW reservation Admission control VM placement BW reservation Admission control TAG spec Network topology & BW reservation state Available VM slots host1 10 host2 50 host3 25 We b (N) DB (N) B

29. Discussion 29

30. Next Time: Centralized Arbitration 30

31. 31