Re: [myrinet] Re: tcp/ip in linux 2.4 kernel

On Thu, 28 Jun 2001, A. Cameron Macdonell wrote:
>
> Here are the CPU usage for messages of size 512 with sending and
> receiver buffers of 32767 by way of the output of top,
>
> I included the CPU monitor stuff as well,
>
> Actually on this test we hit throughput of 1045 Gb/s
>
> Here is the output of top
>
> netperf side:
> CPU0 states:  3.2% user, 59.0% system,  0.0% nice, 37.1% idle
> CPU1 states:  1.4% user, 38.4% system,  0.0% nice, 59.0% idle
>
>   PID USER     PRI  NI  SIZE  RSS SHARE STAT %CPU %MEM   TIME COMMAND
> 13871 cam       15   0   720  720   548 R    99.9  0.1   8:11 netperf
>
>
> netserver side:
> CPU0 states:  0.0% user, 26.3% system,  0.0% nice, 73.1% idle
> CPU1 states:  0.3% user, 67.2% system,  0.0% nice, 31.4% idle
>
>   PID USER     PRI  NI  SIZE  RSS SHARE STAT %CPU %MEM   TIME COMMAND
> 14321 cam       15   0   868  868   436 R    71.1  0.1   0:54 netserver
>
> On Thu, 28 Jun 2001, Patrick Geoffray wrote:


I've never really seen linux get consistently above 1gigabit/sec with
the current GM/ethernet implementation. In fact, when I moved from
using 666MHz processors to 800MHz processors using the exact same
setup, just replacing the CPUs, I saw my tcp performance go down!
There's definitely something fishy with the linux windowing, buffering,
etc. code, or perhaps our driver interacts badly somehow.

1.8Gbit/sec sounds about right for UDP and is a further indication that
linux doesn't have the right sort of flow-control and buffering
strategies for the tcp-side of the IP stack.

------------------------------------------------------------------
Bob Felderman                                 (626) 821-5555
Director of Software Development              (626) 821-5316 fax
Myricom Inc.                                  [email protected]
325 N. Santa Anita Ave.                       http://www.myri.com
Arcadia, CA 91006
------------------------------------------------------------------