[ofiwg] the problems with atomics

Sur, Sayantan sayantan.sur at intel.com
Mon Jul 13 13:31:19 PDT 2015


On 7/13/15, 1:23 PM, "Jason Gunthorpe" <jgunthorpe at obsidianresearch.com>
wrote:

>On Mon, Jul 13, 2015 at 08:20:35PM +0000, Sur, Sayantan wrote:
>> If we do punt the issue to an upper layer, such as MPI, then should we
>> export the fi_datatype_size() method so that apps can find out what
>> libfabric was compiled with?
>
>Isn't it the other way around? libfabric needs to know what the app
>was compiled with?

I assumed that if we punt, then all we need the app to do is to check
whether its definition of long double matched OFI’s, and if not, bail
(which may mean emulate atomics itself).

However, as Jeff Hammond just pointed out, size may not be sufficient
anyways :-|



More information about the ofiwg mailing list