[ofa-general] [PATCH 7/8 V3] core: Add XRC receive-only qp support

Tang, Changqing changquing.tang at hp.com
Tue Feb 5 07:54:54 PST 2008


Thanks, Jack. In High Availability(HA) mode, every rank will register the recvQP, but in non-HA mode, if any rank dies, the whole
application will teardown, so no one register the recvQP, (the creator register it implicitly, right ?)

I have another question about ibv_open_xrc_domain(,fd,). All ranks on a node get the 'fd' to the same unique file, since this is
a temp file, we need to remove it. Can I remove the file right after ibv_open_xrc_domain() returns ? If so, when I start another
MPI job, the same temp file may be generated again for that job, is it possible this job attached to the old job's domain ?

Thanks.

--CQ

> -----Original Message-----
> From: Jack Morgenstein [mailto:jackm at dev.mellanox.co.il]
> Sent: Tuesday, February 05, 2008 12:17 AM
> To: Tang, Changqing
> Cc: Roland Dreier; general at lists.openfabrics.org
> Subject: Re: [ofa-general] [PATCH 7/8 V3] core: Add XRC
> receive-only qp support
>
> On Sunday 03 February 2008 19:30, Tang, Changqing wrote:
> > >
> > > In addition, the patch implements the foundation for distributing
> > > XRC-receive-only QP events to userspace processes registered with
> > > that QP.
> >
> > If r1 is on node1, r2 and r3 are on node2, r1 and r2 have the XRC
> > conection established, r1 knows SRQ number from both r2 and
> r3, but r3
> > does not register the recv QP created by r2, can r3 still receive
> > message from r1 ?
>
> Yes, it can.  However, when r2 either unregisters the recv
> QP, or exits, that QP will be destroyed, and r3 will no
> longer be able to receive messages via the recv QP created by r2.
>
> If, however, r3 does register with the recv QP, that QP will
> not be destroyed until r3 unregisters (or exits).
>
> - Jack
>



More information about the general mailing list