Let It Flow Resilient Asymmetric Load Balancing with Flowlet Switching

Transcription

1 Let It Flow Resilient Asymmetric Load Balancing with Flowlet Switching Erico Vanini*, Rong Pan*, Mohammad Alizadeh, Parvin Taheri*, Tom Edsall* *

2 Load Balancing in Data Centers Multi-rooted tree 1000s of server ports Goal: Avoid congestion hotspots Active research area Solved for symmetric topologies Still open question in asymmetric scenarios 2

3 Asymmetry Is Common in Practice Link Failure 3

4 Asymmetry Is Common in Practice Imbalanced striping: # of ports indivisible by # of switches in other tier 10G$ 2x10G$ S0#######################S1##S2#######################S3##S4######################S5# L0####L1#####################L2####L3#####################L4####L5# Zhou et al. WCMP: Weighted cost multipathing for improved fairness in data centers, CoNEXT

5 Dealing with Asymmetry 25G 50G (TCP) 25G 5

6 Dealing with Asymmetry 20G (UDP) 20G 25G 50G (TCP) 25G 6

7 Dealing with Asymmetry Scheme Thrput 20G (UDP) 20G ECMP-WCMP (Local Stateless) Congestion-Aware 65G 20G 50G (TCP) 25G 7

8 Dealing with Asymmetry Scheme Thrput 20G (UDP) 20G ECMP-WCMP (Local Stateless) Congestion-Aware 65G 70G 15G 50G (TCP) 35G Handling asymmetry needs path congestion information, which varies dynamically with traffic. 8

9 Example: CONGA 1. Leaf switches (top-of-rack) track congestion to other leaves on different paths 2. Use this information to minimize bottleneck utilization L1- L2- Conges'on(To(Leaf- L0 L1 L2 Dest-Leaf- Path

10 Existing Load Balancing Schemes Congestion-aware decisions: complex Measure and feed back congestion in real time CONGA, Hedera, HULA, MPTCP, FlowBender, Congestion-oblivious decisions: simple Random, round robin, hashing decision process ECMP, WCMP, Packet-Spray, Presto, 10

11 Is there a simple load balancing scheme (with congestion-oblivious decisions) that can handle asymmetry? 11

12 LetFlow Simple: Randomly assign Flowlets to available paths Flowlets:.... Δ Δ Δ Flowlets are bursts from same flow separated by at least Δ the main origin of flowlets is the burstiness of TCP at RTT and sub- RTT scales. Kandula et al, Dynamic load balancing without packet reordering, (2007) 12

13 Simple Asymmetric Scenario Detect and randomly assign Flowlets to available paths Link Failure Extremely simple! - No measurements - No feedback - No congestion state 32x10G A 32x10G B 13

14 LetFlow FCT (ms) better ECMP CONGA LetFlow Traffic Load % (w/o link failure) 14

15 What s Going On? Force % of choices per path Spine0 Spine1 Link Failure 32x10G A 32x10G B 15

16 Flowlets are Robust Performance is not sensitive to load balancing decisions Norm. FCT (best) flowlets flows packets (oracle) % of Choices Assigned to Green path Traffic load: 90% 16

17 Flowlet Length avg flowlet length (packets) full capacity path half capacity path Ideal split % of Choices Assigned to Green path 17

18 Flowlets are Elastic Flowlets change size based on congestion on the path - Uncongested path! larger flowlets - Congested path! smaller flowlets! Flowlet sizes implicitly encode path congestion information this determines the amount of traffic on each path not just load balancing decisions LetFlow is congestion-aware, despite simple random decisions 18

19 Why Are Flowlets Elastic? Because of congestion control (e.g., TCP) A flowlet gap occurs on - Window cuts (Loss/ECN) - Latency spikes (ACK clocking) Window halved But, there s a more basic reason, applicable to any congestion control protocol t 19

20 LetFlow Analysis Assume flows transmit as Poisson processes C 1 Avg. rate of each flow: n 1 + n 2 n 1 ƛ 1 = C 1 / n 1 n 2 C 2 ƛ 2 = C 2 / n 2 20

21 LetFlow Analysis n 1 + n 2 n 1 n 2 C 1 C 2...,......,... & State transition probability! "#,"% ( ) %(( +,( - ) /01 ), 4 1,2 21

22 LetFlow Analysis n 1 + n 2 n 1 n 2 C 1 C 2 Takeaways 1. Flows move from low rate paths (small ƛ) to high rate paths (large ƛ)...,......, The flowlet timeout (Δ) is important - Shouldn t be too small or large & State transition probability! "#,"% ( ) %(( +,( - ) /01 ), 4 1,2 22

23 Experiments Summary Different workloads: web search, data mining, enterprise Testbed experiments: ECMP, CONGA, LetFlow 2 leaves 2 spines, 64 servers: symmetric & asymmetric topologies Simulations: ECMP, WCMP, Presto, CONGA, LetFlow Large topology: 6 leaves 6 spines, 288 servers Complex asymmetric topologies: speed mismatch, combined workloads, multitier Different protocols: TCP, DCTCP, DCQCN 23

24 Large Scale Simulations FCT$(ms)$ 500" 400" 300" 200" 100" CONGA" LetFlow" WCMP" Presto" 0" 30" 40" 50" 60" 70" 80" 90" Load$%$$ 10G$ 2x10G$ S0#######################S1##S2#######################S3##S4######################S5# L0####L1#####################L2####L3#####################L4####L5# topology from: WCMP, Zhou, Junlan, et al LetFlow within 2X of CONGA; Both are much better than other schemes 24

25 Multi Destination Scenario spine0 spine1 2x 1x10G 1 : 81 b c 8 : 1 L0 L1 L2 Traffic Load uniform to the 2 destinations 25

26 Multi Destination Scenario spine0 2x 1x10G b spine1 L0 L1 L2 Traffic Load uniform to the 2 destinations c % of destinatioon traffic FCT is similar between Conga and LetFlow CONGA LetFlow 90% LetFlow 60% 26

27 Other Transport Protocols FCT (ms) DCTCP DCQCN ECMP CONGA LetFlow FCT (ms) ECMP CONGA LetFlow Load % (w/o link failure) Load % (w/o link failure) 27

28 Conclusion Flowlet switching is a powerful technique for asymmetric load balancing LetFlow: a simple LB mechanism that handles asymmetry Random decisions but implicitly congestion-aware Suitable for standalone switches does not need feedback Letflow is stochastic and reactive in nature Cannot proactively prevent congestion / queue buildup like more sophisticated schemes 28

29 LET it FLOW! 29