[mvapich-discuss] mvapich2-0.9.8-branch question

[Sundeep Narravula]

Hi Kanoj,

  The RDMA_FAST_PATH that we use in mvapich2 involves polling on the last
byte of an expected message to flag message completions on the remote
side. Since this is not guaranteed by iWARP we have disabled this
by default for the iWARP mode.

Does the RNIC you use guarantee inorder placement of RDMA bytes into the
remote buffer? or guarantee that the last byte will be placed last into
the user buffer?

Regards,
  --Sundeep.


On Wed, 14 Feb 2007, Kanoj Sarcar wrote:

> Hello all,
>
> I am a mpi/mvapich newbie looking at iwarp support in
> mvapich2.
>
> Between 0.9.8 release and branch, in procedure
> rdma_get_control_parameters() in file
> src/mpid/osu_ch3/channels/mrail/src/gen2/ibv_param.c,
> I see a difference that I am trying to understand.
>
> In the release version, the code was:
>
>    if ((value = getenv("MV2_DISABLE_RDMA_FAST_PATH"))
> != NULL) {
>         proc->has_adaptive_fast_path = 0;
>         rdma_polling_set_limit       = 0;
>     } else {
>         proc->has_adaptive_fast_path = 1;
>     }
>
> but this got changed in the branch version to also
> turn off adaptive fast path in case of iwarp.
>
> I am curious about the rationale for this change. Is
> this to mostly support amasso card that (AFAIK) does
> not have user mode fastpath support? Or is there some
> architectural issue in current mvapich2 that requires
> turning off fastpath on iwarp rnics?
>
> Thanks.
>
> Kanoj
>
> PS: please cc me on responses, I am not subscribed to
> the list
>
>
>
> ____________________________________________________________________________________
> Do you Yahoo!?
> Everyone is raving about the all-new Yahoo! Mail beta.
> http://new.mail.yahoo.com
> _______________________________________________
> mvapich-discuss mailing list
> mvapich-discuss at cse.ohio-state.edu
> http://mail.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>