[net-perf] Performance test results

[Michael Sinatra]

Mark Boolootian wrote:
>>Matt was able to install our key into the nms1-losa box as well.  Here
>>are our results:
>>
>>UCB -> Abilene LAX 	971 mb/s
>>Abilene LAX -> UCB	956 mb/s
> 
> 
> We're seeing good performance to this system as well:
> 
>  UCSC -> Abilene LAX    964 mbit/sec
>  Abilene LAX -> UCSC    839 mbit/sec
> 
> Thanks very much for running against the NDT server.  That's the first
> time anyone has reported rates from an off-campus, gigE-connected client.
> 
> 
>>UCB -> nitro.ucsc.edu (NDT)	876 mb/s
>>nitro.ucsc.edu (NDT) -> UCB	882 mb/s
> 
> 
> A bwctl run from nitro to the Berkeley system yielded some interesting
> results:
> 
>   UCSC -> bwctl-001   479 mbit/sec
>   bwctl-001 -> UCSC   945 mbit/sec
> 
> I'd sure like to see roughly the same rates for iperf and the NDT server,
> and that's probably worth looking at a bit more closely.  It looks like
> the MTU on the Berkeley machine isn't set for jumbo frames - though I'm
> puzzled by the reports of different MSS in the below output from iperf.
> The MSS for the Abilene transfer was reported 8140 by both sides.  I'm 
> seeing the differing MSS values to Berkeley whether I'm sending or
> receiving.  I'm also puzzled as to why it reports the MTU on my side as
> 6184, since the interface is set to 8192.

bwctl-001 is connected to the currently-non-jumbo-frame-capable EGM link 
(due to the 4006).  The reason you're seeing "jumbo" frames from ucsc to 
bwctl-001 has to do with the fact that until late this afternoon, 
bwctl-001 and bwctl-002 will continue to be the same box, with the 
default route pointing out the dark fiber link.  This should all make 
much more sense this afternoon when bwctl-001 and bwctl-002 become two 
different boxes and we can test on both.

michael