microsecond shaping issues - cisco by:timevens,#ccie#6109(sp&rs)# email:[email protected]/...

13
An Observation By : Tim Evens, CCIE #6109 (SP & RS) Email: [email protected] / [email protected] Date : 10/2009

Upload: buithu

Post on 29-Apr-2018

217 views

Category:

Documents


3 download

TRANSCRIPT

Page 1: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

An  Observation  

By            :  Tim  Evens,  CCIE  #6109  (SP  &  RS)  Email:  [email protected]  /  [email protected]  Date    :  10/2009  

Page 2: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

Overview  

  Cisco  ME3750’s,  7600  SIP-­‐400/600,  and  ASR1000’s  (not  verified  on  ASR’s  yet)  implement  network  processer  based  shaping  at  the  microsecond  level,  supporMng  ingress  and  egress  shaping.    Shaping  is  performed  using  approximately    a  48us  Tc.  IOS  based  routers,  which  are  preTy  much  all  customer  edge  devices,  support  shaping  egress  at  a  minimum  Tc  of  4ms.  

Page 3: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

  The  Tc  value  matters  because  it  defines  the  burst  window  of  conformance.    Below  illustrates  a  burst  window  of  48us  for  1Gbps  and  100Mbps  

  There  are  20,833  samples/windows  at  48us.    With  a  fixed  Tc,  the  sample/shape  window  remains  the  same,  therefore  the  allowed  burst  shrinks  per  window  as  the  bits  per  second  decreases    

  In  IOS,  Tc  is  derived  by  using  Tc=bc/CIR.    Tc  is  not  configurable  in  IOS  but  Bc  and  CIR  are.    The  lowest  value  for  IOS  is  4ms  

Tc  Matters  

2400

 bits  

1200

 bits  

1200

 bits  

1200

 bits  

2400

 bits  

48us   48us   48us   48us   48us  

4800

 bits  

2400

0bits  

1200

0  bits  

1200

0  bits  

1200

0  bits  

2400

0  bits  

48us   48us   48us   48us   48us  

4800

0  bits  

100M

bps  

1Gbp

s  

48,000  bits    per  Tc   4,800  bits    per  Tc  

Page 4: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

Tc  Compare  

  Using  a  100Mbps  rate,  at  4ms  the  burst  is  400,000  bits  per  shaping  window.  In  other  words;  400,000  bits  *  250  shape  windows  is  100,000,000  bits  per  second.  

  The  Tc  at  48us  gives  us  20,833  shape  windows  with  4800  bits  per  window.  This  isn’t  even  enough  bits  for  a  1500  byte  packet  (12,000  bits).  In  this  case,  a  window  will  be  skipped  in  order  to  credit  enough  bits.  

Page 5: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

1Gbps  to  100Mbps  (shaped)  

  To  calculate  proportions,  we  need  to  bring  values  to  the  same  base  or  interval  in  this  case,  which  is  microseconds,  not  milliseconds.  

  1Gbps  line  rate  is  1000bits  per  microsecond,  resulting  in  48000  bits  per  48us  shape  window.    

  At  100Mbps,  the  allowed  48us  shaped  window  burst  is  4800  bits,  not  48,000  bits.    What  happens  to  the  remaining  43,200  bits?      

  The  answer  is  that  if  Hierarchical  Queuing  Framework  (HQF)  is  used,  the  third  or  second  level  policy  can  queue  the  excess  packets  in  order  to  buffer  long  enough  before  they  they  are  policed/dropped.    If  a  2  or  3  level  queue  is  not  used,  the  packets  will  be  policed  (dropped)  in  the  ME3750  and  7600/SIP-­‐400/600.    

Page 6: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

1Gbps  to  100Mbps  (shaped)  -­‐  Config  

  Below  is  a  ME3750  configuration  example  showing  how  to  queue  the  excess  packets  that  are  shaped  

  policy-­‐map  PE-­‐1G-­‐phy_in    description  -­‐-­‐==|  Physical  policy  –  1Gbps  |==-­‐-­‐                                          class  class-­‐default  

shape  average  1000000000  service-­‐policy  PE-­‐VLAN_shape  

  policy-­‐map  PE-­‐VLAN_shape  Description  -­‐-­‐==|  QoS/shape  at  VLAN  level  |==-­‐-­‐    class  Cust-­‐vlan  

shape  average  100000000                                                          100Mbps  service-­‐policy  PE-­‐QoS_default  

  policy-­‐map  PE-­‐QoS_default  Description  -­‐-­‐==|  QoS/queue  traffic  ***  class  class-­‐default  

bandwidth  remaining  percent  100  queue-­‐limit  9000                                                                                              Enough  packets  to  prevent  or  reduce  drops  

Page 7: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

Problem?  

The  burst  size  from  the  customer  is  4ms  (400,000  bits)  transmitted  at  1Gig  line  rate  which  could  be  transmitted  in  400  microseconds  or  in  bursts  over  the  4ms  window.    As  long  as  the  IOS  4ms  window  is  not  exceeded,  the  shaper  doesn't  queue,  thus  slow  down  the  traffic.    For  this  reason,  microbursts  at  1Gbps  line  rates  slip  through  the  IOS  4ms  shaper  but  run  into  a  brick  wall  on  the  microsecond  shaper  implemented  in  the  ME3750  and  7600  SIP-­‐400/600.      

  The  customer  cannot  reduce  their  egress  traffic  down  to  the  configured  microsecond  shape  burst  window  unless  the  customer  implements  the  same  hardware  with  microsecond  shaping  

  If  the  service  provider  also  shapes  egress  on  the  customer  edge,  then  microbust  convergence  in  the  cloud  is  likely  and  cannot  be  controlled  by  the  customer  unless  symmetric  forwarding  (layer  2  &  3)  or  EVC/VLAN  mapping  are  used  

  Service  providers  normally  charge  extra  for  QoS  and  therefore  will  not  typically  queue/buffer  using  the  third  (second  in  SIP-­‐400/600)  policy,  which  is  required  with  microsecond  shapers  due  to  customer  equipment  restrictions  of  4ms  minimum  shapers  

Page 8: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

Problem  examples  with  Qwest  QMOE  (Layer  2  MPLS/VPLS)  services  

Illustrations  

Page 9: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

Shaping  Tc  Mismatch  Problem  

Packets  are  transmiTed  at  1Gbps  rate,  illustrated  by  the  short    

dash  line  

Packets  are  transmiTed  at    100Mbps  rate,  illustrated  by    

the  long  dash  line  

Qwest  shapes/polices  outgoing  rate  to  purchased  rate  of    

100Mbps.  Packets  at  micro-­‐second      level  are  dropped.  

HUB  

Remote  

Qwest  MAN  Resolu;on  is  to  shape  this  at  the    micro-­‐second  level  to  match  

the  purchased  rate  for  the  desMnaMon    

Page 10: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

SP  Microburst  Convergence  Problem      

HUB  1  Qwest  MAN  

Resolu;on  is  to  remove  one  of  these  flows  or  to  shape  each  ½  the  rate  

so  that  each  flow  does  not  use  more  than  purchased  rate  when  combined.    Or  pay  for  EVC’s  or  QoS  within  Qwest.  

Remote  

Packets  are  transmiTed  at    100Mbps  rate  per  desMnaMon  

purchased  rate  Packets  are  transmiTed  at    100Mbps  rate  per  desMnaMon  purchased  rate  

Packets  converge  to  go  out  to  the  remote    site.  Qwest  switch  sees  both  HUB  site  flows  at  100Mbps  each,  combined  rate  is  200Mbps.  Qwest  drops  packets  egress  due  to  the  over    

subscribed  rate  of  200Mbps.      

Rate  is  100Mbps  and  contains  both  HUB  site  flows.  

HUB  2  

Page 11: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

Shared  VLAN  Microburst  Multicast  Problem  

HUB  1  

Remote  1  

Qwest  MAN  

Resolu;on  is  to  forward  this  mulMcast  traffic  via  unicast  over  the    Qwest  shared  vlan  or  purchase  EVC’s    

where  Qwest  maps  VLANS  to  specific  sites.    

HUB  2  

This  remote  requests  to    receive  mulMcast  from  HUB  

site  1.  

Remote  2  

HUB  1  sends  mulMcast  to  remote  site  1  as  requested.  

Remote  site  2  receives  remote  site  1’s  mulMcast  because  of  the  

 shared  Qwest  VLAN.  

HUB  2  receives  remote  site  1’s  mulMcast  because  of  the  

 shared  Qwest  VLAN.  

Page 12: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

Conclusion  

Page 13: Microsecond Shaping Issues - Cisco By:TimEvens,#CCIE#6109(SP&RS)# Email:tim@evensweb.com/ tim.evens@swedish.org# Date:10/2009#

Conclusion  

Cisco  introduced  ME3750  and  7600  SIP-­‐400/600  devices  that  support  microsecond  shaping.    Customers  are  still  using  hardware  that  only  supports  a  minimum  of  4ms  shaping.    

Due  to  the  microsecond  shaping,  the  customer  traffic  at  4ms  shaping  can/will  microburst  greater  than  the  48us  shaper  window,  resulting  in  packet  drops  in  the  service  provider  network.    

The  customer  is  left  hanging  and  either  forced  to  purchase  new  hardware  to  terminate  the  service  provider  connections,  EVC’s  to  map  layer  2  VLANS,  or  QoS  to  buffer  the  packets  that  would  have  been  dropped  by  the  microsecond  shaper.  

If  microsecond  shapers  are  used  by  service  providers,  then  service  providers  should  by  default  buffer  customer  traffic  enough  to  handle  the  microburst  up  to  4ms.