SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter - PowerPoint PPT Presentation

SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks Long Luo Hongfang Yu Klaus-Tycho Foerster Stefan Schmid Faculty of Computer Science University of Electronic Science University of Vienna and Technology of China Austria P.R. China UESTC China

Multicast workload in datacenter network Modern datacenter applications are rife with point-to-multipoint master communication patterns---multicast workloads workers Data analytics applications Database query workers master Pub/Sub workers Publish-subscribe systems Iterative machine learning jobs SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 2

Hybrid datacenter network Kl ToR ToR Packet-switched inter-rack network A A circuit switch B B C C D D The circuit switch can build directed port-to-port or port-to-multiport circuit connections between the ToRs. SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 3

Hybrid datacenter network Kl ToR ToR ToR ToR Packet-switched Packet-switched network network A A A A circuit switch circuit switch B B B B C reconfigure C C C D D D D The circuit switch can be reconfigured to change circuit connections between the ToRs. SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 4

Intuition • Physical layer multicasting improves the performance in transferring multicast flows ⎯ packets can be delivered to multiple ToRs in a single transmission ⎯ high-bandwidth up to 40GbE or 100GbE • Physical layer multicasting > IP unicasting • Physical layer multicasting > IP multicasting SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 5

Challenges • How to schedule multicast flows efficiently? • fully use the network bandwidth SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 6

Intuition #1 • unit flows f 1 , f 2 , f 3 , each transfer data from one rack to another two port capacity =1 f 1 R 1 f 2 R 2 f 3 R 3 average flow time: !"#"$ !" $ SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 7

Intuition #1 • unit flows f 1 , f 2 , f 3 , each transfer data from one rack to another two port capacity =1 f 1 f 1 R 1 R 1 f 23 f 2 R 2 f 3 R 2 f 21 R 3 f 3 R 3 split f 2 into f 21 and f 23 average flow time: !"#"$ !" $ average flow time: !"#"# ≈$.&' $ SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 8

Intuition #1 • unit flows f 1 , f 2 , f 3 , each transfer data from one rack to another two port capacity =1 f 1 f 1 R 1 R 1 f 23 f 2 R 2 f 3 R 2 f 21 R 3 f 3 R 3 split f 2 into f 21 and f 23 average flow time: !"#"$ !" $ average flow time: !"#"# ≈$.&' $ un splittable multicast < splittable multicast SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 9

Insights into the fundamental static problem • The ( unsplittable ) multicast matching problem • Equivalent to a specific hypergraph matching problem • NP-hard even for k = 2 receivers per transfer • If each source has at most one transfer: • Polynomial-time for k = 2 receivers per transfer • NP-hard for every k > 2 SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 10

Insights into the fundamental static problem • The ( unsplittable ) multicast matching problem • Equivalent to a specific hypergraph matching problem • NP-hard even for k = 2 receivers per transfer • If each source has at most one transfer: • Polynomial-time for k = 2 receivers per transfer When considering splittable case: • Polynomial-time for any k • NP-hard for every k > 2 SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 11

Multicast scheduling problem • Objectives • Maximizing the network throughput • Minimizing the flow time • Which circuit connections should be configured? • When to preempt flows and reconfigure circuit connections? SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 12

Intuition #2 sender- time needs #Flows receivers to complete splittable multicast + non-preemptive scheduling #ports f 1 1→2,3,4 20ms f 1 101 122 142 1 1 21 f 2 5→1,3,6 20ms reconfiguration f 3 2 f 1 f 3 3→1,2,5 20ms f 4 3 f 2 f 5 f 4 1→5,6 100ms 4 f 3 f 5 3→1,2,4 40ms 5 f 4 f 2 f 2 6 f 5 Time(ms) T 1 =[1ms, 101ms] T 2 =[102ms, 142ms] SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 13

Intuition #2 sender- time needs #Flows receivers to complete splittable multicast + non-preemptive scheduling #ports f 1 1→2,3,4 20ms f 1 101 122 142 1 1 21 f 2 5→1,3,6 20ms reconfiguration f 3 2 f 1 f 3 3→1,2,5 20ms f 4 3 f 2 f 5 f 4 1→5,6 100ms 4 f 3 f 5 3→1,2,4 40ms 5 f 4 f 2 f 2 6 f 5 Time(ms) T 1 =[1ms, 101ms] T 2 =[102ms, 142ms] splittable multicast + preemptive scheduling #ports f 1 1 1 21 42 63 123 f 4 reconfiguration f 3 2 f 1 f 4 3 f 2 f 5 f 5 4 f 3 5 improve average flow time by 1.74 × f 4 f 2 f 2 6 f 5 T 1 =[1ms, 21ms] T 2 =[22ms, 42ms] T 3 =[43ms, 123ms] Time(ms) SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 14

Intuition #2 splittable multicast > unsplittable multicast preemptive scheduling > non-preemptive scheduling SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 15

Solution #1 • Formulate as an optimization problem Constraints • Circuit switch port: each port can be involved in one connect • Link and port capacities • Flow sizes Maximizing the network throughput Minimizing the flow time SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 16

Solution #1 • Algorithm design • Hierarchically creating circuit connections and scheduling flows • Calculating the epoch length to maximize the network throughput or minimize the flow time SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 17

Challenge #2 • Receiver asynchronization An example of transferring three units of data to three receivers S d 1 d 2 d 3 1 st epoch SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 18

Challenge #2 • Receiver asynchronization An example of transferring three units of data to three receivers S d 1 d 2 d 3 1 st epoch 2 nd epoch SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 19

Challenge #2 • Receiver asynchronization An example of transferring three units of data to three receivers S d 1 d 2 d 3 1 st epoch 2 nd epoch 3 rd epoch SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 20

Challenge #2 • Receiver asynchronization An example of transferring three units of data to three receivers S d 1 ✘ d 2 ✘ d 3 ✓ 1 st epoch 2 nd epoch 3 rd epoch 4 th epoch SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 21

Solution #2 • Receiver asynchronization An example of transferring three units of data to three receivers reconfigure three times 4 th epoch 5 th epoch 6 th epoch flow time: 3+3 𝛆 duration: 1 duration: 1 duration: 1 SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 22

Solution #2 • Receiver asynchronization An example of transferring three units of data to three receivers reconfigure three times 4 th epoch 5 th epoch 6 th epoch additional flow time: 3+3 𝛆 duration: 1 duration: 1 duration: 1 tradeoff between reconfiguration and reconfigure once the amount of data sent to network additional flow time: 3+ 𝛆 4 th epoch duration: 3 SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 23

Evaluation • Comparison • Blast: non-preemptive scheduling + unsplittable multicast Xia, Yiting, et. al. "Blast: Accelerating high-performance data analytics applications by optical multicast." 2015 INFOCOM. • Creek: preemptive scheduling + unsplittable multicast Sun, Xiaoye Steven, et.al. "When creek meets river: Exploiting high-bandwidth circuit switch in scheduling multicast data." 2017 ICNP • Splitcast: preemptive scheduling + splittable mutlicast SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 24

Evaluation • Splitcast vs. Creek vs. Blast (c) (a) (b) (d) SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 25

Summary and Outlook • We exploit enablers of reconfigurable datacenter networks: • in-network multicast • splittable multicasting • preemptive scheduling • simulations show good performance in flow time and throughput • Outlook: • find and test further realistic workloads • Extend to multi-hop routing SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 26

Thanks! SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks (INFOCOM20) 27