32
Ankit Singla Data center architecture

Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

  • Upload
    others

  • View
    8

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

Ankit SinglaData center architecture

Page 2: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

2

A server rack

A rack of servers

A top-of-rack switch

Page 3: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

3

Lots of racks

How to network the racks?

Page 4: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

“Big switch” approach

4

Page 5: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

“Big switch” approach

5

Page 6: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

Jupiter Rising: A Decade of Clos Topologies and

Centralized Control in Google’s Datacenter Network

Arjun Singh, Joon Ong, Amit Agarwal, Glen Anderson, Ashby Armistead, Roy Bannon,Seb Boving, Gaurav Desai, Bob Felderman, Paulie Germano, Anand Kanagala, Jeff Provost,Jason Simmons, Eiichi Tanda, Jim Wanderer, Urs Hölzle, Stephen Stuart, and Amin Vahdat

Google, Inc.j it i @ l

ACM SIGCOMM, 2015

512 ports x 1G

40 servers x 1G

4 ports x 1G

610:1 “oversubscription”

Page 7: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

“Big switch” approach

7

$$$$$$$$$$ $$$$$$$$$$

Page 8: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

$$$ $$$

Alternative: tree network

8

Page 9: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

9

Connect many cheap, identical switches?

Goals: high capacity, low latency

Page 10: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

If you know your application …

10

Page 11: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

… design for it

11

NERSC

Page 12: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

But, other apps may not work well …

12

Page 13: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

13

What’s so hard about this?6-port switches

4-port switches

���

�����

�����

�����

������

������

� � � �� �� �� �� �� ��

������������������������

������ �� ��������

Page 14: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

14

What’s so hard about this?

Page 15: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

15

Clos networks

Use small, cheap elements to build large networks!!

Page 16: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

Fat-tree

Page 17: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

17

Fat-tree network

Pod 0

10.0.2.1

10.0.1.1

Pod 1 Pod 3Pod 210.2.0.2 10.2.0.3

10.2.0.1

10.4.1.1 10.4.1.2 10.4.2.1 10.4.2.2

10.2.2.1

10.0.1.2

ACM SIGCOMM, 2008

A Scalable, Commodity Data Center Network Architecture

Mohammad Al-Fares Alexander Loukissas Amin Vahdat

Servers

Core

Page 18: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

18

Fat-tree network

Page 19: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

19

Jupiter Rising: A Decade of Clos Topologies and

Centralized Control in Google’s Datacenter Network

Arjun Singh, Joon Ong, Amit Agarwal, Glen Anderson, Ashby Armistead, Roy Bannon,Seb Boving, Gaurav Desai, Bob Felderman, Paulie Germano, Anand Kanagala, Jeff Provost,Jason Simmons, Eiichi Tanda, Jim Wanderer, Urs Hölzle, Stephen Stuart, and Amin Vahdat

Google, Inc.j it i @ l

ACM SIGCOMM, 2015

Page 20: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

[Image: Robert Harker]

Page 21: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

[Image: Robert Harker]

Page 22: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

22

Jupiter Rising: A Decade of Clos Topologies and

Centralized Control in Google’s Datacenter Network

Arjun Singh, Joon Ong, Amit Agarwal, Glen Anderson, Ashby Armistead, Roy Bannon,Seb Boving, Gaurav Desai, Bob Felderman, Paulie Germano, Anand Kanagala, Jeff Provost,Jason Simmons, Eiichi Tanda, Jim Wanderer, Urs Hölzle, Stephen Stuart, and Amin Vahdat

Google, Inc.j it i @ l

ACM SIGCOMM, 2015

Page 23: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

23

Cabling[Robert Harker

Page 24: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

24

DC architecture

How to network the racks?

But can we do better?

Page 25: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

25

Other topologies besides fat-trees?

6-port switches

4-port switches

���

�����

����������������������

� � � �� �� �� �� �� ��������������������������

������ �� ��������

Surprise: picking randomly from this space works better than fat-trees!

Deterministic expander constructions can also achieve these gains

Page 26: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

But do we need full bandwidth everywhere at all times?

Page 27: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

25% is the high mark for overall utilization (?)  ”

  2012.

  Facebook

 

 

27

Page 28: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

hotspots

Skew in network traffic

Page 29: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

29

How does this observation help?

“Unfortunately, given current data center network architectures, the only way to provision required bandwidth bandwidth between dynamically changing sets of nodes is to build a non-blocking switch fabric at the scale of an entire data center”

.

Page 30: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

Set up network connections on the fly?

Servers

ToR

switch

Aggregate

switch

Core

switch

Electrical

Network

Optical

Network

Reconfigurable

optical paths

Array of Micromirrors

Diffracted beam Towards destinationReceived beam

Input beam LasersDMDs

Photodetectors

Mirror assembly Reflected beam

TX

RX

Rack%1%Rack%N%Rack%r%

Steerable%%FSOs%

Ceiling%mirror%

ToR%switch%

FireFly%Controller%

Traffic%Pa=erns%

Rule%change%

FSO%reconf%

Can statically wired networks compete with this?

Page 31: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,

31

DC architecture @ SIGCOMM 2017RotorNet: A Scalable, Low-complexity,

Optical Datacenter Network

William M. Mellette, Rob McGuinness, Arjun Roy, Alex Forencich,

George Papen, Alex C.Snoeren, and George Porter

University of California, San Diego

A

T

a

l

e

o

f

T

w

o

T

o

p

o

l

o

g

i

e

s

:

E

x

p

l

o

r

i

n

g

C

o

n

v

e

r

t

i

b

l

e

D

a

t

a

C

e

n

t

e

r

N

e

t

w

o

r

k

A

r

c

h

i

t

e

c

t

u

r

e

s

w

i

t

h

F

l

a

t

-

t

r

e

e

Yiting Xia, Xiaoye Steven Sun, Simbarashe Dzinamarira,

Dingming Wu, Xin Sunny Huang, T. S. Eugene Ng

Rice University

Beyond fat-trees without antennae, mirrors, and disco-balls

Simon KassingETH Zürich

[email protected]

Asaf ValadarskyHebrew University of Jerusalem

[email protected]

Gal ShahafHebrew University of Jerusalem

[email protected]

Michael SchapiraHebrew University of Jerusalem

[email protected]

Ankit SinglaETH Zürich

[email protected]

Page 32: Data center architecture - acm sigcomm · 2017-10-27 · Jupiter Rising: A Decade of Clos Topologies and Centralized Control in Google’s Datacenter Network Arjun Singh, Joon Ong,