[j-nsp] Juniper's Gigabit performance

[Josef Buchsteiner]

Wednesday, December 24, 2003, 9:22:30 PM, you wrote:
> On Dec 24, 2003, at 9:10 AM, Josef Buchsteiner wrote:
>>
>>   There  is  no J-Cell going from the PIC towards the FPC so this
>>   is  net.  The  packet  stream  will be cut into Cell at the I/O
>>   Manager  towards  Memory  and this is where you have the memory
>>   bandwidth of 3.2Gbps excluding J-Cell overhead for one I/O ASIC
>>   and  the  reason  why you will not be able to run line rate on
>>   4 * 1port GE-PIC on one FPC which is also documented AFAIK.
>>
> I think there are some big misconceptions people have.  Does this mean
> that if you have 4 gig-e's doing random packet len, 0bps in one
> direction, and 1000mbps in the other, it would max out at an aggregate
> of ~3.2gbps or ~2.5gbps of actual IP data?


  let   me   explain  more in detail to clear things up. The memory
  bandwidth  from  the  I/O Manager is 32bit @125Mhz. which gives
  you a raw bandwidth of 4Gbps. Since each J-Cell has an overhead
  of  16 byte we come to the 3.2Gbps of J-Cell payload. From here
  on  we only talk about cells per second. So we can have at max.
  per  FPC  an  aggregate  throughput  of  4Gbps/((64*8)+(16*8))=
  6,25Mcps.

  So  the question is IP-Payload identical to J-Cell Payload. Not
  all  the  time  and  it  depends on the packet-size. Lets do en
  example.

  64byte  Ethernet Frame is 1488095,2 pps @46IP. So if you have a
  stream  of 4 PIC's and each sending this amount we need 1488096
  cells  *  4 = 5952384 Cells per second in total. So that should
  work.  Lets  do  an example of 1500 byte. Here we need 24 Cells
  and  with 1500byte at IP we can send max 81274,382 pps. So we need
  81275  *  24  = 1950600 cells per second for one Stream. If you
  multiply  this  by four you would need 7802400 cells/per second
  which  is 20% more what we can handle. ( I've not mentioned for
  simplicity  that  we  add  some more J-Cells without ip-payload
  which  contain  address  information only for larger packets ).
  Now you can do the academic maths. for all packets sizes.


  hope this clarifies the topic
  Josef
  
  
  


> --Phil Rosenthal
> ISPrime, Inc.

> _______________________________________________
> juniper-nsp mailing list juniper-nsp at puck.nether.net
> http://puck.nether.net/mailman/listinfo/juniper-nsp