[ofa-general] RDMA write completion question

[Michael Krause]

At 07:34 AM 5/25/2007, Caitlin Bestler wrote:

>Steve Wise Wrote:
>
>-----Original Message-----
>
>Devesh Sharma wrote:
> > On 5/24/07, Gleb Natapov <glebn at voltaire.com> wrote:
> >> Hi,
> >>
> >>  Does local RDMA write completion guaranties that a data that was
> >> RDMAed is
> >> already accessible in a destination's host _memory_?
> > Local RDMA write completion guarantees that the data you have RDMAed
> > has been copied into the remote buffer, without any data corruption.
> >>
>For iWARP, the local write completion simply means you can reuse the
>local buffer and the the transport will deliver it or kill the
>connection.  The data _could_ be queued in the local rnic and anywhere
>else in the tcp cloud.
>
>
>_______________________________________________
>
>
>And The only real difference with InfiniBand is that the uncertainty
>cloud is limited to the gap between the HCA and the application.
>Protocol designers can debate the tradeoffs InfiniBand takes to
>achieve that, but the import thing to Application Designers is
>that "smaller" is not "zero".
>
>Generally, once a Send has been posted that all other interactions
>with the remote peer over the same connection can assume that the prior
>actions have completed, but if your application needs an absolute
>guarantee that something has happened (for checkpointing or other
>purposes) then you really can only rely on a peer-to-peer message.

The peer-to-peer may be either a RDMA Read which is not visible to the 
application / ULP or may be an application / ULP interaction.   As a 
general guideline, application developers should not assume anything about 
the end-to-end data integrity or delivery being guaranteed by the hardware 
and take appropriate steps to design their communication pattern to 
validate the data was correctly exchanged.  This is not difficult and is 
often built into many ULP or applications already.

Mike





>_______________________________________________
>general mailing list
>general at lists.openfabrics.org
>http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
>
>To unsubscribe, please visit 
>http://openib.org/mailman/listinfo/openib-general