[ofiwg] strided IOVs

[Underwood, Keith D]

Simple strides are a fairly specific use case that apply to dense matrices and regular problems.  I am not denying that those problems exist, but they are a subset.  As I understand it, regular applications are a shrinking fraction of the application domain.

Does any current hardware support strides natively?  If not, maybe we have a little time to work out that interface.

From: Jeff Hammond [mailto:jeff.science at gmail.com]
Sent: Thursday, October 29, 2015 3:09 PM
To: Underwood, Keith D <keith.d.underwood at intel.com>
Cc: Hefty, Sean <sean.hefty at intel.com>; Howard Pritchard <hppritcha at gmail.com>; ofiwg at lists.openfabrics.org
Subject: Re: [ofiwg] strided IOVs

Both GA/ARMCI and SHMEM utilize regular aka simple strided operations in RMA.  GA_Acc/ARMCI_AccS/MPI_Accumulate map to fi_atomic, that means those also need to support strided, assuming RMA does.

I think only simple strided is worth supporting, because it is O(1) metadata.  General gather/scatter ala iovec is unlikely to optimizable relative to a higher-level software implementation, so I don't see a good reason to put it into OFI.

Jeff

On Thu, Oct 29, 2015 at 4:46 AM, Underwood, Keith D <keith.d.underwood at intel.com<mailto:keith.d.underwood at intel.com>> wrote:
Exactly.  However, we should be careful about what we are adding here.  A strided interface is a very specific special case that applies to regular problems.  There is a bready of MPI derived datatype capabilities that are not covered by a simple stride.  What are the objectives for covering datatypes?  Is the IOVEC interface the right place to glue in the first handling of a special case of datatypes?

> -----Original Message-----
> From: ofiwg-bounces at lists.openfabrics.org<mailto:ofiwg-bounces at lists.openfabrics.org> [mailto:ofiwg-<mailto:ofiwg->
> bounces at lists.openfabrics.org<mailto:bounces at lists.openfabrics.org>] On Behalf Of Hefty, Sean
> Sent: Wednesday, October 28, 2015 3:14 PM
> To: Hefty, Sean <sean.hefty at intel.com<mailto:sean.hefty at intel.com>>; Howard Pritchard
> <hppritcha at gmail.com<mailto:hppritcha at gmail.com>>
> Cc: ofiwg at lists.openfabrics.org<mailto:ofiwg at lists.openfabrics.org>
> Subject: Re: [ofiwg] strided IOVs
>
> > I need more context to understand how writing X bytes every N bytes
> > into a one-time use receive buffer is useful.
>
> FWIW, someone sent me this link:
>
> http://people.csail.mit.edu/fred/ghost_cell.pdf
>
> as an example.
> _______________________________________________
> ofiwg mailing list
> ofiwg at lists.openfabrics.org<mailto:ofiwg at lists.openfabrics.org>
> http://lists.openfabrics.org/mailman/listinfo/ofiwg
_______________________________________________
ofiwg mailing list
ofiwg at lists.openfabrics.org<mailto:ofiwg at lists.openfabrics.org>
http://lists.openfabrics.org/mailman/listinfo/ofiwg



--
Jeff Hammond
jeff.science at gmail.com<mailto:jeff.science at gmail.com>
http://jeffhammond.github.io/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/ofiwg/attachments/20151108/aa77b412/attachment.html>