[ofiwg] [libfabric-users] RDM send fails
sayantan.sur at intel.com
Thu Nov 29 09:29:48 PST 2018
Wanted to add that in general an RDM send can EAGAIN (not fail), for any provider dependent reason. So ... in general the apps would have to guard against a send call returning EAGAIN.
From: Libfabric-users <libfabric-users-bounces at lists.openfabrics.org> on behalf of Jörn Schumacher <jorn.schumacher at cern.ch>
Date: Thursday, November 29, 2018 at 9:23 AM
To: "Hefty, Sean" <sean.hefty at intel.com>, "libfabric-users at lists.openfabrics.org" <libfabric-users at lists.openfabrics.org>
Cc: "ofiwg at lists.openfabrics.org" <ofiwg at lists.openfabrics.org>
Subject: Re: [libfabric-users] RDM send fails
On 11/29/2018 05:59 PM, Hefty, Sean wrote:
>> Thanks Sean. Couldn't the connection be established already when
>> inserting the remote address to the AV? From the API it looks like an
>> event can be generated once the insertion completes. In case of rxm
>> this could be when the connection is established.
> I thought about this, but the addresses can be inserted prior to the AV being bound to any EP, plus the AV can be bound to multiple EPs. So there may be more than one connection per peer address that gets established, and we may not know the source address to use for the connection.
> There are some options here, but we'd likely need to add a flag as part of AV insertion to force the connection setup to complete, plus require that the AV be bound to all EPs prior to the insert being called. Neither of these are hard, but apps would need to adjust for this case.
Got it. Sounds like it might break some compatibility then. Not sure if
it's worth it for my particular use case, unless such a change is of
Libfabric-users mailing list
Libfabric-users at lists.openfabrics.org
More information about the ofiwg