• No results found

CS 91: Cloud Systems & Datacenter Networks Networks Background

N/A
N/A
Protected

Academic year: 2021

Share "CS 91: Cloud Systems & Datacenter Networks Networks Background"

Copied!
36
0
0

Loading.... (view fulltext now)

Full text

(1)

CS  91:  Cloud  Systems  &  

Datacenter  Networks  

 

(2)
(3)

Agenda  

• 

Overview  of  tradiBonal  network  topologies  

• 

IntroducBon  to  soEware-­‐defined  networks  

(4)

Topology  

• 

Shape  and  structure  of  the  network  

“Star”  or  

“Hub  and  spoke”   “Bus”  

(5)

Topology  

• 

Shape  and  structure  of  the  network  

“Tree”  

Most  common  in  

datacenter.    Why?  

(6)

Growth  

“Star”  or  

(7)

Growth  

(8)

“TradiBonal”  Datacenter  

40  Servers   40  Servers   40  Servers   40  Servers   Edge  

(9)

“TradiBonal”  Datacenter  

40  Servers   40  Servers   40  Servers   40  Servers   Edge  

(10)

“TradiBonal”  Datacenter  

40  Servers   40  Servers   40  Servers   40  Servers   Edge  

AggregaBon   Core  

(11)

Recall:  OversubscripBon  

40  Servers   40  Servers   40  Servers   40  Servers   Edge  

AggregaBon   Core  

(12)

OversubscripBon  

40  Servers   40  Servers   40  Servers   40  Servers   Edge   AggregaBon   Core  

All  hosts  

connected  to  

same  switch.    

Full  speed  

between  all  of  

them.    (1:1)  

(e.g.,  1  Gbps)  

(13)

OversubscripBon  

40  Servers   40  Servers   40  Servers   40  Servers   Edge  

AggregaBon   Core  

If  these  links  are  

also  1  Gbps,  we  

now  have  1/40

th

 of  

link  capacity  for  

each  host.      (40:1)  

What  might  we  do  to  

help  alleviate  this?  

(14)

OversubscripBon  

40  Servers   40  Servers   40  Servers   40  Servers   Edge  

AggregaBon   Core  

10  Gbps  gives  us  

a  raBo  of  4:1.  

Good  enough?  

Highlighted  sub-­‐secBon  might  be  OK  if:  

• 

ApplicaBon  has  good  locality  

(15)

OversubscripBon  

40  Servers   40  Servers   40  Servers   40  Servers   Edge  

AggregaBon   Core  

Might  wants  hosts  to  

have  10  Gbps  too!  

SBll  problemaBc.  

(16)

OversubscripBon  

40  Servers   40  Servers   40  Servers   40  Servers   Edge  

AggregaBon   Core  

(17)

Redundancy  (MulB-­‐rooted  Tree)  

40  Servers   40  Servers   40  Servers   40  Servers   Edge  

AggregaBon   Core  

(18)

Redundancy  (MulB-­‐rooted  Tree)  

40  Servers   40  Servers   40  Servers   40  Servers   Edge  

AggregaBon   Core  

Pro:  More  capacity!  Con:  Decision  must  be  made  (complexity,  reordering)  

1  

?  

(19)

Equal-­‐Cost  MulB  Path  (ECMP)  

• 

If  there’s  a  decision  to  be  made,  hash  packet  

• 

All  packets  of  flow  (conversaBon)  use  same  path  

• 

Hash  collisions  lead  to  poor  uBlizaBon  

(20)

Paper  Preview  

• 

“A  Scalable,  Commodity  Data  Center  Network  

Architecture”  

– 

AlternaBve  tree  redundant  tree  topology  design  

• 

“Hedera:  Dynamic  Flow  Scheduling  for  Data  

Center  Networks”  

– 

Use  live  traffic  informaBon  to  decide  path  choice  

• 

“AugmenBng  Data  Center  Networks  with  

(21)

Network  Hardware  

Tradi&onal  Hardware  

SDN  Hardware  

You  Can’t  read  this,  it’s   too  small!    Seriously,   stop  trying  to  read  this,   more  important  things   are  happening  

elsewhere!  

You  Can’t  read  this,  it’s   too  small!    Seriously,   stop  trying  to  read  this,   more  important  things   are  happening  

elsewhere!   You  Can’t  read  this,  it’s  

too  small!    Seriously,   stop  trying  to  read  this,   more  important  things   are  happening  

elsewhere!  

LeE   Right  

This  sounds  great,  why  mess  with  a  good  thing?  

 

(22)

SoEware-­‐Defined  Networking  (SDN)  

Tradi&onal  Hardware  

SDN  Hardware  

Controller  

You  Can’t  read  this,  it’s   too  small!    Seriously,   stop  trying  to  read  this,   more  important  things   are  happening  

elsewhere!   You  Can’t  read  this,  it’s  

too  small!    Seriously,   stop  trying  to  read  this,   more  important  things   are  happening  

elsewhere!  

LeE  @  100  Mbps   Right  @  300  Mbps  LeE  @  400  Mbps   Right  @  200  Mbps  

(23)

Centralized  Network  Control  

• 

Not  so  great  on  the  Internet  

– 

Failures  common  

– 

Long  distances  

– 

Who  would  be  in  charge?  

• 

Powerful  in  controlled  environment  

(24)

Centralized  Network  Control  

• 

Treat  device  as  a  blank  forwarding  table.  

Protocol   Src   Dst   Src_Pt   Dst_Pt   Output  

(25)

Centralized  Network  Control  

• 

Table  built  from  TCAM  memory.  

Protocol   Src   Dst   Src_Pt   Dst_Pt   Output  

TCP   10.1.1.1   10.1.1.2   54321   80   1   *   10.1.3.3   10.1.3.4   *   *   2  

(26)

Centralized  Network  Control  

• 

Table  built  from  TCAM  memory.  

Protocol   Src   Dst   Src_Pt   Dst_Pt   Output  

TCP   10.1.1.1   10.1.1.2   54321   80   1   *   10.1.3.3   10.1.3.4   *   *   2  

(27)

Paper  (&  Lab)  Preview  

• 

“Ethane:  Taking  Control  of  the  Enterprise”  

• 

“OpenFlow:  Enabling  InnovaBon  in  Campus  

Networks”  

• 

“Extending  Networking  into  the  VirtualizaBon  

Layer”  

• 

Lab:  Using  SDN  (OpenFlow)  to  implement  

(28)

Layering  

• 

Primary  network  abstracBon:  stack  of  layers  

• 

Each  layer  has  disBnct  responsibiliBes  

(29)

OSI  Seven-­‐Layer  Model    

ApplicaBon:  the  applicaBon  (e.g.,  the  Web,  Email)  

Transport:  end-­‐to-­‐end  connecBons,  reliability   Network:  rouBng  

Link  (data-­‐link):  framing,  error  detecBon  

Physical:  1’s  and  0’s/bits  across  a  medium  (copper,  the  air,  fiber)   PresentaBon:  formanng,  encoding,  encrypBon  

(30)

OSI  Seven-­‐Layer  Model    

ApplicaBon:  the  applicaBon  (e.g.,  the  Web,  Email)  

Transport:  TCP  /  UDP   Network:  IP  

Link  (data-­‐link):  Ethernet   Physical:  Wires  

“Layer  7”  

“Layer  3”  

(31)

Header  EncapsulaBon  

ApplicaBon  Data

 

Transport:  TCP

 

Network:  IP

  (payload)  

Link:  Ethernet

  (payload)  

(payload)  

(32)

Header  EncapsulaBon  

ApplicaBon  Data

 

Transport:  TCP

 

Network:  IP

  (payload)  

Link:  Ethernet

  (payload)  

(payload)  

This  whole  chunk  of  bytes  is  what  goes  out  on  the  wire.  

Adds  reliability  informaBon.  

Adds  port  numbers  to  

(33)

Header  EncapsulaBon  

ApplicaBon  Data

 

Transport:  TCP

 

Network:  IP

  (payload)  

Link:  Ethernet

  (payload)  

(payload)  

This  whole  chunk  of  bytes  is  what  goes  out  on  the  wire.  

Adds  addresses.  

Network:  IP  address  

Link:  MAC  address  

(34)

Addressing  

• 

Layer  3  /  Network  /  IP  addresses:  

– 

Hierarchical  rouBng  

– 

The  numerical  values  say  something  about  the  

locaBon  of  the  node  in  the  network  

• 

Layer  2  /  Data  Link  /  MAC  addresses:  

– 

Just  a  flat,  unique  idenBfier  

(35)

“Middlebox”  

• 

(Usually)  transparent  device  that  looks  at  

traffic  passing  through,  someBmes  modifies  it  

• 

Examples:  

– 

Firewall:  Look  at  traffic  headers.    If  the  “wrong”  

machines  are  communicaBng,  block  it.  

(36)

Other  terms  

• 

Please  post  on  Piazza  or  ask  in  class  if  you  

References

Related documents

And NovaStor NovaBAKCUP® DataCenter and xSP offer enterprise-level cross-platform backup and restore for local networks, private cloud networks and data

(Sulphur Dioxide returns to picket line. Nitros walk over.)?. Harry: (to the audience) He's

The first part comprised of selective questions which was aimed to point various primitive aspects of electronic governance implementation including the basis of

efficiencies and other characteristic values were very low, the power generations were confirmed in all solid state solar cells fabricated in this study using Fe 2 O 3 as the

Furthermore, this complex communication between pseu- dogenes, miRNAs, and mRNA transcripts affects the valida- tion of putative miRNA targets, given that certain cells may have

It has been well understood that the performance of a cloud application heavily depends on its allocated share of resources, including computing resources, such as CPU and memory,

We investigated whether MSCs migrate to OA tissues and if this migration could be influenced by adapting the culture conditions before application of MSCs. Both OA synovium

Covid 19 Travel Industry Research Data Provider.. ROI Measurement