[mvapich-discuss] Is a construct like this able to take advantage of GDR?

[Subramoni, Hari]

Hi, Raghu.

As long the format is MPI compliant, such a call can be used to initiate GPU to GPU transfers with MVAPICH2-GDR.

Thx,
Hari.

From: mvapich-discuss-bounces at cse.ohio-state.edu On Behalf Of Raghu Reddy
Sent: Tuesday, February 13, 2018 3:54 PM
To: mvapich-discuss at cse.ohio-state.edu <mvapich-discuss at mailman.cse.ohio-state.edu>
Subject: [mvapich-discuss] Is a construct like this able to take advantage of GDR?

A user has the following construct in the code that he is trying to benchmark:

--------------------------------
!$acc host_data use_device(sedgvarRecv)

CALL MPI_IRECV(sedgvarRecv(:,:,:,:,n,varNumber), sedgvarSize, MPI_INTEGER, ...)
---------------------------------

I don't know if this is sufficient information provide the necessary context to answer the question, but I was concerned that the array syntax in the argument would cause problems because MPI would be using  temporary variable to pass the argument.

Anyone has any comments if this is a valid construct for testing data transfers from GPU to GPU?

Thanks,
Raghu

-------------- next part --------------
A non-text attachment was scrubbed...
Name: winmail.dat
Type: application/ms-tnef
Size: 7042 bytes
Desc: not available
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20180213/70dbcac7/attachment.bin>