Pr Prog ogrammin ramming g th the e T op opolo ology gy of - - PowerPoint PPT Presentation

pr prog ogrammin ramming g th the e t op opolo ology gy
SMART_READER_LITE
LIVE PREVIEW

Pr Prog ogrammin ramming g th the e T op opolo ology gy of - - PowerPoint PPT Presentation

Pr Prog ogrammin ramming g th the e T op opolo ology gy of N f Netw etworks orks T ec echno hnolo logy gy and nd Alg lgor orithms ithms Manya Ghoba badi di Pierre Blanche Klaus-Tycho Foerster Daniel Kilper


slide-1
SLIDE 1

Pr Prog

  • grammin

ramming g th the e T

  • p
  • polo
  • logy

gy of N f Netw etworks

  • rks

T ec echno hnolo logy gy and nd Alg lgor

  • rithms

ithms

Pierre Blanche Klaus-Tycho Foerster Jeff Cox Jamie Gaudette Nikhil Devanur Phillipa Gill Mark Filer Madeleine Glick Daniel Kilper Gireeja Ranade Janar nardhan dhan Kulkar karni i Houman Rastegarfar Ratul Mahajan Stefan Schmid Amar Phanishayee Rachee Singh

Manya Ghoba badi di

slide-2
SLIDE 2

Th The e clou

  • ud

d infras frastructu tructure

Inefficiencies and waste in the cloud infrastructure

Data centers Optical cables

2

slide-3
SLIDE 3

Network topology Traffic engineering

Capacity provisioning T ec echn hnolo

  • logy

gy and d algor gorithms ithms to to op

  • pti

timize mize net etwor

  • rk to

topol

  • logy
  • gy

3

slide-4
SLIDE 4

High gh-lev level el ide dea Networks with programmable topologies A B D C

1

A B D C

1 1 1 1

Throughput: 2 units Throughput: 3 units

1 1 1

Dynamic topology Static topology

4

slide-5
SLIDE 5
  • Challenging

llenging: :

  • Requires reconfigurable hardware te

technolo chnology gy

  • Requires revisiting networking layer algori

rithms thms

  • Imp

mpactful ctful:

  • Cheaper networks
  • Higher throughput

Pr Prog

  • gramm

rammab able le to topol

  • logie
  • gies

5

slide-6
SLIDE 6

T alk lk ou

  • utl

tline ine

Data ta cente nter netw tworks

  • rks

ProjecT

  • R: Programming the

network topology [SIGCOMM’16]

T echnology and algorithms to enable programmable topologies in the cloud

Google data center

Wide de-area networks tworks

Programming the capacity of links

[SIGCOMM’18, HotNets’17, OFC’16]

Level3 global backbone

6

slide-7
SLIDE 7

T

  • day’s data center interconnects

A B C D

3 3 3 3 3 3 3 3 3 3 3 3

Ideal demand matrix: uniform and static Static capacity between T

  • p-of-Rack (T
  • R) pairs

6 6

Non-ideal demand matrix: skewed and dynamic A B C D A B C D A B C D A B C D

10Gbps 10Gbps 8 6 7 7 12 8 6 6

7

slide-8
SLIDE 8

Nee eed for

  • r a r

a rec econ

  • nfigurab

figurable le inter erconne connect ct

Data:

  • 200K servers across 4 production clusters
  • Cluster sizes: 100 -- 2500 racks

Observation:

  • Many rack pairs exchange little traffic
  • Only some hot rack pairs are active

Implication:

  • Static topology with uniform capacity:
  • Over-provisioned for most rack pairs
  • Under-provisioned for few others

Reconfigurable interconnect: T

  • dynamically provide additional capacity between hot rack pairs

8

slide-9
SLIDE 9

Ou Our r propo

  • posal:

sal: Pr Proj

  • jec

ecT

  • R
  • R inter

tercon connect nect

9

  • Free-space topology (programmable)
  • Digital micromirror device to redirect light
  • Disco-ball shaped mirror assembly to magnify reachability

Laser Photodetector Static topology

9

slide-10
SLIDE 10

Di Digital gital Micr cromirr

  • mirror
  • r De

Device ce (DMD) DMD)

Array of micromirrors (10 um) Memory cell

10

slide-11
SLIDE 11

A 3-T

  • R
  • R Pr

Proj

  • jec

ecT

  • R
  • R inte

terconnect connect pr prot

  • tot
  • type

ype

T

  • R1

T

  • R2

T

  • R3

Source laser

DMD

Mirrors reflecting to T

  • R2 and T
  • R3

T

  • R1

T

  • R2

T

  • R3

11

slide-12
SLIDE 12

Rou

  • uting

ting algor gorithm ithm

  • We have a highly flexible topology allowing for millions of ways to

connect lasers to photodetectors

  • Ideal solution: fast changing topology to adapt to demand change
  • Challenge: It takes 12μs to reprogram a link

ToR1 ToR1 ToR2 ToR3 ToR2 ToR3

lasers photodetectors

12

slide-13
SLIDE 13

Rou

  • uting

ting algor gorithm ithm

  • Two topology approach:
  • Slow switching topology or dedicated topology
  • Fast switching links or opportunistic links

ToR1 ToR2 ToR3 ToR1 ToR2 ToR3 ToR1 ToR1 T

  • R2

ToR3 ToR2 ToR3 ToR1 ToR2 T

  • R3

ToR1 ToR2 T

  • R3

lasers photodetectors

dedicated topology

  • pportunistic links

13

slide-14
SLIDE 14

Rou

  • uting

ting packets ets

ToR1 ToR2 ToR3 ToR1 ToR2 ToR3

dedicated topology

2

3 3 3

Virtual output queues

K-shortest paths routing

  • pportunistic link

2 2 2 2 2 2 2

14

slide-15
SLIDE 15

Scheduling eduling op

  • ppor
  • rtun

tunistic istic lin inks ks

  • Given a set of potential links and current traffic demand, find a set
  • f active opportunistic links

ToR1 ToR2 ToR3 ToR1 T

  • R2

ToR3

100 100 s

  • u

r c e

d e s t i n a t i o n

15

slide-16
SLIDE 16

Scheduling eduling op

  • ppor
  • rtun

tunistic istic lin inks ks

  • Standard switch scheduling problem
  • Blossom matching
  • BvN matrix decomposition
  • Centralized scheduler
  • Single tiered matching

100 100 100 s

  • u

r c e

d e s t i n a t i o n

input

  • utput

16

slide-17
SLIDE 17
  • Standard switch scheduling problem
  • Blossom matching
  • BvN matrix decomposition
  • Centralized scheduler
  • Single tiered matching

Scheduling eduling op

  • ppor
  • rtun

tunistic istic lin inks ks

input

  • utput

Src T

  • Rs

Dst T

  • Rs

Two-tiered Decentralized Extended the Gale-Shapely algorithm for finding stable matches [GS-1962] Constant competitive against an offline optimal allocation

17

slide-18
SLIDE 18
  • Slow switching time

+ Reconfigurable + Switching time: 12μs

  • T

ail flow completion time

  • Different traffic matrices
  • Impact of switching time
  • Impact of fan-out

Simu mulation lation res esul ults ts

5 10 15 20 25 30 35 40

20 30 40 50 60 70 80

Average Flow Completion Time (ms)

Average Load (%)

  • No reconfigurability

95%

ProjecT

  • R

Fat tree

[SIGCOMM’08]

FireFly [SIGCOMM’14]

18

slide-19
SLIDE 19

Th The e key ey ta takea eaway way from

  • m th

this is ta talk Current assumption: Network topology is fixed New world: Network topology is dynamic

c3 c4

s w u v y t

c7 c8 Problems to solve:

  • Scheduling
  • Capacity provisioning
  • Traffic engineering
  • Load-balancing
  • Exci

citing ting: Unusual wealth of algorithms

  • Challe

lleng ngin ing: : Changes fundamental assumptions

  • Imp

mpactfu ful: l: Better efficiency ($/Gbps)

19