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

▶
splitcast optimizing multicast flows in reconfigurable
SMART_READER_LITE
LIVE PREVIEW

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

Nov 03, 2022 220 likes •505 views

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

slide-1
SLIDE 1

SplitCast: Optimizing Multicast Flows in Reconfigurable Datacenter Networks

Long Luo Hongfang Yu Klaus-Tycho Foerster Stefan Schmid

University of Electronic Science and Technology of China P.R. China Faculty of Computer Science University of Vienna Austria

UESTC China

slide-2
SLIDE 2

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

Multicast workload in datacenter network

2

Data analytics applications Modern datacenter applications are rife with point-to-multipoint communication patterns---multicast workloads

Iterative machine learning jobs master workers workers master workers

Publish-subscribe systems

Pub/Sub

Database query

slide-3
SLIDE 3

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

Hybrid datacenter network

3

A B C D ToR

Packet-switched inter-rack network circuit switch

A B C D ToR Kl The circuit switch can build directed port-to-port or port-to-multiport circuit connections between the ToRs.

slide-4
SLIDE 4

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

Hybrid datacenter network

4

reconfigure A B C D ToR

Packet-switched network circuit switch

A B C D ToR A B C D ToR

Packet-switched network circuit switch

A B C D ToR Kl The circuit switch can be reconfigured to change circuit connections between the ToRs.

slide-5
SLIDE 5

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

Intuition

5

  • 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
slide-6
SLIDE 6

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

Challenges

  • How to schedule multicast flows efficiently?
  • fully use the network bandwidth

6

slide-7
SLIDE 7

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

Intuition #1

7

  • unit flows f1, f2, f3, each transfer data from one rack to another two

f2 f3 R1 R2 R3 f1 average flow time: !"#"$

$

!"

port capacity =1

slide-8
SLIDE 8

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

Intuition #1

8

f1 f23 f21 f3 R1 R2 R3

  • unit flows f1, f2, f3, each transfer data from one rack to another two

f2 f3 R1 R2 R3 f1 split f2 into f21 and f23 average flow time: !"#"#

$

≈$.&' average flow time: !"#"$

$

!"

port capacity =1

slide-9
SLIDE 9

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

Intuition #1

9

  • unit flows f1, f2, f3, each transfer data from one rack to another two

f2 f3 R1 R2 R3 f1

unsplittable multicast < splittable multicast

average flow time: !"#"$

$

!"

port capacity =1

f1 f23 f21 f3 R1 R2 R3 split f2 into f21 and f23 average flow time: !"#"#

$

≈$.&'

slide-10
SLIDE 10

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

Insights into the fundamental static problem

10

  • 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
slide-11
SLIDE 11

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

Insights into the fundamental static problem

11

  • 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

When considering splittable case:

  • Polynomial-time for any k
slide-12
SLIDE 12

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

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?

12

slide-13
SLIDE 13

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

Intuition #2

13

f1 f4

T1=[1ms, 101ms] T2=[102ms, 142ms]

f2 f3 f5 f2

1 2 3 4 5 6 #ports

splittable multicast + non-preemptive scheduling

21 101 122 reconfiguration Time(ms) 1

f2 f1 f3 f4 f5

142

#Flows sender- receivers

time needs to complete

f1 1→2,3,4 20ms f2 5→1,3,6 20ms f3 3→1,2,5 20ms f4 1→5,6 100ms f5 3→1,2,4 40ms

slide-14
SLIDE 14

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

Intuition #2

14

21 42 63 123 reconfiguration Time(ms) 1

f2 f1 f3 f4 f5 splittable multicast + preemptive scheduling

#ports

f1 f4 f3 f5

T1=[1ms, 21ms] T2=[22ms, 42ms]

f4

T3=[43ms, 123ms]

f5 f2 f2

1 2 3 4 5 6

#Flows sender- receivers

time needs to complete

f1 1→2,3,4 20ms f2 5→1,3,6 20ms f3 3→1,2,5 20ms f4 1→5,6 100ms f5 3→1,2,4 40ms

f1 f4

T1=[1ms, 101ms] T2=[102ms, 142ms]

f2 f3 f5 f2

1 2 3 4 5 6 #ports

splittable multicast + non-preemptive scheduling

21 101 122 reconfiguration Time(ms) 1

f2 f1 f3 f4 f5

142 improve average flow time by 1.74×

slide-15
SLIDE 15

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

Intuition #2

15

splittable multicast > unsplittable multicast preemptive scheduling > non-preemptive scheduling

slide-16
SLIDE 16

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

Solution #1

  • Formulate as an optimization problem

16

Maximizing the network throughput Minimizing the flow time Constraints

  • Circuit switch port: each port can be involved in one connect
  • Link and port capacities
  • Flow sizes
slide-17
SLIDE 17

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

Solution #1

  • Algorithm design
  • Hierarchically creating circuit connections and scheduling flows
  • Calculating the epoch length to maximize the network throughput
  • r minimize the flow time

17

slide-18
SLIDE 18

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

Challenge #2

  • Receiver asynchronization

18

An example of transferring three units of data to three receivers

S d1 d2 d3

1st epoch

slide-19
SLIDE 19

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

Challenge #2

  • Receiver asynchronization

19

An example of transferring three units of data to three receivers

S d1 d2 d3

1st epoch 2nd epoch

slide-20
SLIDE 20

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

Challenge #2

  • Receiver asynchronization

20

An example of transferring three units of data to three receivers

S d1 d2 d3

1st epoch 2nd epoch 3rd epoch

slide-21
SLIDE 21

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

Challenge #2

  • Receiver asynchronization

21

An example of transferring three units of data to three receivers

4th epoch

S d1 d2 d3

1st epoch 2nd epoch 3rd epoch

✘ ✘ ✓

slide-22
SLIDE 22

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

Solution #2

  • Receiver asynchronization

22

An example of transferring three units of data to three receivers

4th epoch 5th epoch 6th epoch duration: 1 duration: 1 duration: 1

reconfigure three times flow time: 3+3𝛆

slide-23
SLIDE 23

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

Solution #2

  • Receiver asynchronization

23

An example of transferring three units of data to three receivers

4th epoch 5th epoch 6th epoch 4th epoch

tradeoff between reconfiguration and the amount of data sent to network

duration: 1 duration: 1 duration: 1 duration: 3

reconfigure three times reconfigure once additional flow time: 3+𝛆 additional flow time: 3+3𝛆

slide-24
SLIDE 24

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

Evaluation

  • Comparison
  • Blast: non-preemptive scheduling + unsplittable multicast
  • Creek: preemptive scheduling + unsplittable multicast
  • Splitcast: preemptive scheduling + splittable mutlicast

24

Xia, Yiting, et. al. "Blast: Accelerating high-performance data analytics applications by optical multicast." 2015 INFOCOM. Sun, Xiaoye Steven, et.al. "When creek meets river: Exploiting high-bandwidth circuit switch in scheduling multicast data." 2017 ICNP

slide-25
SLIDE 25

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

Evaluation

  • Splitcast vs. Creek vs. Blast

25

(a) (b) (c) (d)

slide-26
SLIDE 26

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

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

26

slide-27
SLIDE 27

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

Thanks!

27