[openib-general] RC v/s UD

[dotanb at dev.mellanox.co.il]

if you want to check performance i suggest you to check the ib_*_lat
ib_*_bw tests and not the pingpong tests which were written as an examples
..

Dotan

>  One more thing: Did you use the default message size of the tests?
>>    In ibv_ud_pingpong the default message size is 2K
>>    In ibv_rc_pingpong the default message size is 4K
>>
>> so, 2 posts and 2 completions where handled in UD for every 1 post and 1
>> completion in RC ...
>
>
>
> I used default message size. ibv_rc_pingpong with message size set to 2K
> gives same reading as ibv_ud_pingpong and with increasing message size
> gives better results.
>
> So does it mean that more number of posts and completions hurt the
> perforrance?? Is there a way to minimize number of posts/completions in UD
> ??
>
> send_bw shows that UD and RC give almost same performance when n = 1000
> iterations but for few iterations (say 2) UD is good.
>
> Basically I am doing some experiments with Broadcast and my readings show
> that for large data sizes the performance is not good. Given that switch
> is
> able to route at high rate, I think the reason for low performance boils
> down to UD being not able to handle more then 2K message size. Is it
> possible to have something like RDMA for UD and hence for
> broadcast (although IB spec does not suppot it) or have hardware do the
> fragmentation for UD if size specified is more then 2K ??
>
> Regards,
> John T.
>