Home > SIGCOMM Preview Session:
 Data Center Networking (DCN)

SIGCOMM Preview Session:
 Data Center Networking (DCN)

Page 1
SIGCOMM  Preview  Session:
 Data  Center  Networking  (DCN)
George  Porter,  UC  San  Diego   2015
These  slides  are  licensed  under  a  Creative  Commons     Attribution-��NonCommercial-��ShareAlike  4.0  International  license

Page 2
��The  cloud��
2

Page 3
��The  cloud��
2

Page 4
��The  Cloud��  =  Lots  of  computing  and  data
Data
+ =

Page 5
��The  Cloud��  =  Lots  of  computing  and  data
Data
+ =

Page 6
��The  Cloud��  =  Lots  of  computing  and  data
Data
+ =
App 1

Page 7
��The  Cloud��  =  Lots  of  computing  and  data
Data
+ =
App 1 App 2

Page 8
��The  Cloud��  =  Lots  of  computing  and  data
Data
+ =
App 1 App 3 App 2

Page 9
��The  Cloud��  =  Lots  of  computing  and  data
Data
+ =
App 1 App 3 App ... App 2 App ... App ... App ... App ... App ... App ... App ... App ... App ...

Page 10
��The  Cloud��  =  Lots  of  computing  and  data
Data
+ =
App 1 App 3 App ... App 2 App ... App ... App ... App ... App ... App ... App ... App ... App ...

Page 11
Computing  and  data  has  to  live  somewhere��
Microsoft
Google Facebook Microsoft

Page 12
Inside  a  data  center

Page 13
Inside  a  data  center
• 10s  or  100s  of  thousands  of  servers

Page 14
Inside  a  data  center
• 10s  or  100s  of  thousands  of  servers • Petabytes  of  data  storage

Page 15
Inside  a  data  center
• 10s  or  100s  of  thousands  of  servers • Petabytes  of  data  storage • Single  ��applications��  spread  across  many   thousands  of  servers  (e.g.,  Amazon.com)  
– Application  components  such  as  caches,  web   servers,  data  bases,  distributed  file  servers,  ��   – Each  component  is  ��scaled��  to  meet  needs  of   millions  of  users

Page 16
Why  study  DCNs?
6

Page 17
Why  study  DCNs?
• Scale  
– Google:  0  to  1B  users  in  ~15  years   – Facebook:  0  to  1B  users  in  ~10  years   – Must  operate  at  the  scale  of  O(1M+)  users
6

Page 18
Why  study  DCNs?
• Scale  
– Google:  0  to  1B  users  in  ~15  years   – Facebook:  0  to  1B  users  in  ~10  years   – Must  operate  at  the  scale  of  O(1M+)  users
• Cost:  
– To  build:  Google  ($3B/year),  MSFT  ($15B/total)   – To  operate:  1-��2%  of  global  energy  consumption*   – Must  deliver  apps  using  efficient  HW/SW  footprint
6
*  LBNL,  2013.

Page 19
What  defines  a  data  center  network?
The  Internet Data  Center  Network  (DCN)
Many  autonomous  systems  (ASes) One  administrative  domain Distributed  control/routing Centralized  control  and  route  selection Single  shortest-��path  routing Many  paths  from  source  to  destination Hard  to  measure Easy  to  measure,  but  lots  of  data�� Standardized  transport  (TCP  and  UDP) Many  transports  (DCTCP,  pFabric,  ��) Innovation  requires  consensus  (IETF) Single  company  can  innovate ��Network  of  networks�� ��Backplane  of  giant  supercomputer��

Page 20
What  defines  a  data  center  network?
The  Internet Data  Center  Network  (DCN)
Many  autonomous  systems  (ASes) One  administrative  domain Distributed  control/routing Centralized  control  and  route  selection Single  shortest-��path  routing Many  paths  from  source  to  destination Hard  to  measure Easy  to  measure,  but  lots  of  data�� Standardized  transport  (TCP  and  UDP) Many  transports  (DCTCP,  pFabric,  ��) Innovation  requires  consensus  (IETF) Single  company  can  innovate ��Network  of  networks�� ��Backplane  of  giant  supercomputer��

Page 21
What  defines  a  data  center  network?
The  Internet Data  Center  Network  (DCN)
Many  autonomous  systems  (ASes) One  administrative  domain Distributed  control/routing Centralized  control  and  route  selection Single  shortest-��path  routing Many  paths  from  source  to  destination Hard  to  measure Easy  to  measure,  but  lots  of  data�� Standardized  transport  (TCP  and  UDP) Many  transports  (DCTCP,  pFabric,  ��) Innovation  requires  consensus  (IETF) Single  company  can  innovate ��Network  of  networks�� ��Backplane  of  giant  supercomputer��

Page 22
What  defines  a  data  center  network?
The  Internet Data  Center  Network  (DCN)
Many  autonomous  systems  (ASes) One  administrative  domain Distributed  control/routing Centralized  control  and  route  selection Single  shortest-��path  routing Many  paths  from  source  to  destination Hard  to  measure Easy  to  measure,  but  lots  of  data�� Standardized  transport  (TCP  and  UDP) Many  transports  (DCTCP,  pFabric,  ��) Innovation  requires  consensus  (IETF) Single  company  can  innovate ��Network  of  networks�� ��Backplane  of  giant  supercomputer��

Page 23
What  defines  a  data  center  network?
The  Internet Data  Center  Network  (DCN)
Many  autonomous  systems  (ASes) One  administrative  domain Distributed  control/routing Centralized  control  and  route  selection Single  shortest-��path  routing Many  paths  from  source  to  destination Hard  to  measure Easy  to  measure,  but  lots  of  data�� Standardized  transport  (TCP  and  UDP) Many  transports  (DCTCP,  pFabric,  ��) Innovation  requires  consensus  (IETF) Single  company  can  innovate ��Network  of  networks�� ��Backplane  of  giant  supercomputer��

Page 24
What  defines  a  data  center  network?
The  Internet Data  Center  Network  (DCN)
Many  autonomous  systems  (ASes) One  administrative  domain Distributed  control/routing Centralized  control  and  route  selection Single  shortest-��path  routing Many  paths  from  source  to  destination Hard  to  measure Easy  to  measure,  but  lots  of  data�� Standardized  transport  (TCP  and  UDP) Many  transports  (DCTCP,  pFabric,  ��) Innovation  requires  consensus  (IETF) Single  company  can  innovate ��Network  of  networks�� ��Backplane  of  giant  supercomputer��

Page 25
What  defines  a  data  center  network?
The  Internet Data  Center  Network  (DCN)
Many  autonomous  systems  (ASes) One  administrative  domain Distributed  control/routing Centralized  control  and  route  selection Single  shortest-��path  routing Many  paths  from  source  to  destination Hard  to  measure Easy  to  measure,  but  lots  of  data�� Standardized  transport  (TCP  and  UDP) Many  transports  (DCTCP,  pFabric,  ��) Innovation  requires  consensus  (IETF) Single  company  can  innovate ��Network  of  networks�� ��Backplane  of  giant  supercomputer��

Page 26
What  defines  a  data  center  network?
The  Internet Data  Center  Network  (DCN)
Many  autonomous  systems  (ASes) One  administrative  domain Distributed  control/routing Centralized  control  and  route  selection Single  shortest-��path  routing Many  paths  from  source  to  destination Hard  to  measure Easy  to  measure,  but  lots  of  data�� Standardized  transport  (TCP  and  UDP) Many  transports  (DCTCP,  pFabric,  ��) Innovation  requires  consensus  (IETF) Single  company  can  innovate ��Network  of  networks�� ��Backplane  of  giant  supercomputer��

Page 27
DCN  research  ��cheat  sheet��

Page 28
DCN  research  ��cheat  sheet��
• How  would  you  design  a  network  to  support   1M  endpoints?

Page 29
DCN  research  ��cheat  sheet��
• How  would  you  design  a  network  to  support   1M  endpoints? • If  you  could��  
– Control  all  the  endpoints  and  the  network?   – Violate  layering,  end-��to-��end  principle?   – Build  custom  hardware?   – Assume  common  OS,  dataplane  functions?

Page 30
DCN  research  ��cheat  sheet��
• How  would  you  design  a  network  to  support   1M  endpoints? • If  you  could��  
– Control  all  the  endpoints  and  the  network?   – Violate  layering,  end-��to-��end  principle?   – Build  custom  hardware?   – Assume  common  OS,  dataplane  functions?
Top-��to-��bottom  rethinking  of  the  network

Page 31
Paper  previews:  Topologies

Page 32
Tree-��based  network  topologies
10

Page 33
Tree-��based  network  topologies
10
100,000  x  10  Gb/s  =  1  Pb/s

Page 34
Tree-��based  network  topologies
10
100,000  x  10  Gb/s  =  1  Pb/s
Can��t  buy   sufficiently  fast   core  switches!

Page 35
11
Folded-��Clos  multi-��rooted  trees

Page 36
11
Al  Fares,  et  al.,   Sigcomm��08
Folded-��Clos  multi-��rooted  trees

Page 37
11
10  Gb/s   servers Al  Fares,  et  al.,   Sigcomm��08
Folded-��Clos  multi-��rooted  trees

Page 38
11
10  Gb/s   Switches 10  Gb/s   servers Al  Fares,  et  al.,   Sigcomm��08
Folded-��Clos  multi-��rooted  trees

Page 39
11
10  Gb/s   Switches 10  Gb/s   servers Al  Fares,  et  al.,   Sigcomm��08
Bandwidth  needs  met  by  massive  multipathing
Folded-��Clos  multi-��rooted  trees

Page 40
Paper  previews:  Topologies
Jupiter  Rising:  A  Decade  of  Clos  Topologies  and   Centralized  Control  in  Google��s  Datacenter  Network   (Singh  et  al.)  
– Tu  5pm-��6:15pm  Session  3.2:  Experience  Track:  2   – 10  year  retrospective  on  Google��s  experiences  building  large-�� scale  networks  
Condor:  Better  Topologies  through  Declarative  Design   (Schlinker  et  al.)  
– Th  8:50am  -��  10:30am  Session  8:  Data  center  networking   – Describing  and  reasoning  about  the  network  structure

Page 41
Paper  previews:  Measurement

Page 42
Network  measurement

Page 43
Network  measurement
• Measuring  the  Internet:  
– No  central  vantage  point,  only  indirect  access  to   certain  portions,  multiple  ASes  hiding  information��

Page 44
Network  measurement
• Measuring  the  Internet:  
– No  central  vantage  point,  only  indirect  access  to   certain  portions,  multiple  ASes  hiding  information��
• Measuring  data  centers:  
– Need  low  latency   – Need  fine-��grained  precision  (milli-��  or  microsecond)   – An  enormous  amount  of  data  to  collect   – Hard  to  publish  findings  (proprietary  data  sets)

Page 45
Paper  previews:  Measurement  (1/2)
Inside  the  Social  Network's  (Datacenter)  Network   (Roy  et  al.)  
– Tu  4pm-��4:50pm  Session  3.1:  Experience  Track  1   – Measurement  study  of  Facebook��s  data  center  
Pingmesh:  A  Large-��Scale  System  for  Data  Center   Network  Latency  Measurement  and  Analysis  (Guo   et  al.)  
– Tu  4pm-��4:50pm  Session  3.1:  Experience  Track  1   – Experience  paper  on  Microsoft��s  system  for  collecting   inter-��server  ping  times  at  scale

Page 46
Paper  previews:  Measurement  (2/2)
Packet-��Level  Telemetry  in  Large  Datacenter   Networks  (Zhu  et  al.)  
– Th  8:50am  -��  10:30am  Session  8:  Data  center   networking   – Packet  tracing  system  deployed  at  Microsoft   designed  for  finding  network  faults

Page 47
Paper  previews:
 Packet/flow  handling

Page 48
Packet  and  flow  handling

Page 49
Packet  and  flow  handling
• Internet  service  model:  
– Best-��effort,  ��end-��to-��end  principle��,  generally  just   one  path  to  a  destination

Page 50
Packet  and  flow  handling
• Internet  service  model:  
– Best-��effort,  ��end-��to-��end  principle��,  generally  just   one  path  to  a  destination
• Data  center  networks:  
Load  balancing:  how  to  effectively  use  all  the  many   paths  to  a  given  destination?   – Better  than  best-��effort:  how  to  prioritize,  rate-��limit,   adjust  relative  sending  rates��

Page 51
Paper  previews:  Packet/flow  handling
Presto:  Edge-��based  Load  Balancing  for  Fast   Datacenter  Networks  (He  et  al.)  
– Th  8:50am  -��  10:30am  Session  8:  Data  center  networking   – Choosing  paths  for  packets  with  help  from  endhosts  
Enabling  End-��Host  Network  Functions  (Ballani  et  al.)  
– Th  8:50am  -��  10:30am  Session  8:  Data  center  networking   – Providing  better  than  best-��effort  handling  of  packets  with   help  from  endhosts

Page 52
In  closing
• DCN  is  an  exciting,  fun  research  area   • While  many  papers  are  from  Microsoft,   Google,  Facebook,  ��  
– YOU  have  the  ability  to  have  enormous  impact   – Many  projects  are  open-��source  
• E.g.,  http://opencompute.org  
• Rethink  the  entire  network  stack!  
– Hardware,  software,  protocols,  OS,  NIC,  ��

Set Home | Add to Favorites

All Rights Reserved Powered by Free Document Search and Download

Copyright © 2011
This site does not host pdf,doc,ppt,xls,rtf,txt files all document are the property of their respective owners. complaint#nuokui.com
TOP