Gleb Natapov wrote:
On Wed, Sep 13, 2006 at 07:50:02PM +0530, Sunil Patil wrote:
Hi,
This is somewhat irrelevant question. It was said that in order delivery is
a feature of Mellanox HCA and not part of IB spec. Is this true for UD type
QPs also for which IB spec says that "in order delivery" is not guaranteed ?
UD is unreliable datagram, Nothing is guaranteed about it. The Mellanox HCA
guarantees inorder delivery of RDMAed data.
InfniBand specification guarantees no placement ordering, but delivery
ordering (messages are delivered in the order sent) for reliable
transport. For un-reliable, since some packets could be lost in the
fabric, those packet will not be delivered, thus no ordering guaranteed.
Mellanox HCAs guarantee in-order placement between packets in reliable
transports, as well as in-packet data placement (last byte of the packet
payload will be written last to the memory).
For un-reliable transport, the ordering between packets can not be
guaranteed (again, since some packets could be lost in the fabric) but
the in-packet placement ordering is guaranteed.
Regards
Sunil
On 9/13/06, Open MPI <b...@open-mpi.org> wrote:
#366: OpenIB on IA64: we should disable small message RDMA
-------------------------+--------------------------------------------------
Reporter: jsquyres | Owner: gleb
Type: defect | Status: new
Priority: major | Milestone: Open MPI 1.1.2
Version: trunk | Resolution:
Keywords: openib, ia64 |
-------------------------+--------------------------------------------------
Comment (by gleb):
Actually in order delivery is a feature of Mellanox HCA and not part of
the IB/iWarp spec.[[BR]]
See [http://openib.org/pipermail/openib-general/2006-August/025305.html]
for discussion about this. So use_eager_rdma flag is a good candidate to
be put in mca-btl-openib-hca-params.ini file.
--
Ticket URL: <http://svn.open-mpi.org/trac/ompi/ticket/366>
Open MPI <http://www.open-mpi.org/>
_______________________________________________
bugs mailing list
b...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/bugs
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
--
Gleb.
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
