[c-nsp] 7200 / NPE-G2

Mark Taylor maillist at smashie.ision.co.uk
Wed Apr 25 11:15:28 EDT 2007


In case this is useful to anyone... We're replacing some G1's with G2's in 
our network at the moment and had an initial concern where the CPU didn't 
drop at all on low loaded boxes after the upgrade. We decided to run some 
tests whilst trying to put some realistic higher load through them and got 
the following between 2 GigE interfaces (one of them dot1q trunking):

5 minute input rate 745653000 bits/sec, 178808 packets/sec
5 minute output rate 431726000 bits/sec, 127011 packets/sec

CPU utilization for five seconds: 82%/81%; one minute: 82%; five minutes: 
82%

router1#sh ip cache flow
IP packet size distribution (2052M total packets):
   1-32   64   96  128  160  192  224  256  288  320  352  384  416  448 
480
   .000 .026 .107 .000 .390 .000 .000 .000 .000 .000 .000 .000 .000 .000 
.000

    512  544  576 1024 1536 2048 2560 3072 3584 4096 4608
   .000 .000 .000 .404 .071 .000 .000 .000 .000 .000 .000

We're just using them as base L3 devices (BGP, OSPF, HSRP, Dot1q, Netflow). 
There aren't any advanced features like long ACLs, QOS, L2TP, NBAR.

Our G1's run at 55% CPU for 250Mbps each way @40Kpps. Pushing 70-80% CPU 
(about as far as we wanted to go) we expected them to get to 350Mbps 
@60Kpps. There's no way the G1 would get near the input/output rate and pps 
rates that we achieved in the testing (in our scenario at least). We're 
aiming for 700Mbps @120Kpps on the G2's and safe working load.

It's a bit daunting rolling out higher spec engines and then finding the CPU 
remains the same. After the tests (and this thread) we're pretty confident 
that as you push them you'll get more from them, so we're not as worried as 
we were before.

Regards,
Mark.


> Thanks Euan.
>
> Note that the pps number really should be doubles because it
> was bidirectional traffic. IIRC it was 64 byte packets.
>
> So the G1 showed some dropped frames at 800kpps and the
> G2 showed some drops at 1.5Mpps. This was a quick test we
> did to simply compare the CPU numbers with the load increase.
>
> Rodney
>
> On Wed, Apr 25, 2007 at 02:26:03PM +0100, Euan Galloway wrote:
>> On Wed, Apr 25, 2007 at 08:49:53AM -0400, Rodney Dunn wrote:
>> > Here is a rough spreadsheet with the numbers we had done.
>>
>> For the attachment challenged amoungst us...
>>
>> (excuse the bad reformatting).
>> (CPU) (CPU)
>> Pps NPE-G1 NPE-G2
>> each direction
>>
>> 100,000 31% 49%
>> 125,000 37% 51%
>> 150,000 44% 55%
>> 175,000 53% 58%
>> 200,000 61% 60%
>> 225,000 67% 61%
>> 250,000 75% 63%
>> 275,000 84% 65%
>> 300,000 90% 68%
>> 325,000 98% 70%
>> 350,000 100% 73%
>> 375,000 100% 75%
>> 400,000 dropped packets 400,000 pps 78%
>> 425,000 81%
>> 450,000 83%
>> 475,000 86%
>> 500,000 88%
>> 525,000 91%
>> 550,000 94%
>> 575,000 95%
>> 600,000 97%
>> 625,000 98%
>> 650,000 99%
>> 675,000 99%
>> 700,000 99%
>> 725,000 99%
>> 750,000 dropped packets at 750,000 pps



More information about the cisco-nsp mailing list