[OFIWG-MPI] [ofiwg] Call today

Hefty, Sean sean.hefty at intel.com
Thu May 22 12:29:22 PDT 2014


Thanks, Howard, this is helpful.

Regarding the 'tag match class' that you mention, would you create an 'rma class' as a peer, with the RMA operations defined in a similar fashion?  If not, why not?  Would this also extend to all other data transfer operations?  I.e. message queue (send/receive) and atomics, plus any others defined in the future?


> -----Original Message-----
> From: Howard Pritchard [mailto:hppritcha at gmail.com]
> Sent: Thursday, May 22, 2014 11:53 AM
> To: Richard Graham
> Cc: Hefty, Sean; ofiwg at lists.openfabrics.org; ofiwg-
> mpi at lists.openfabrics.org
> Subject: Re: [ofiwg] Call today
> 
> Hi Folks,
> 
> here is a diagram of a concept that was discussed in a side conversation at
> the last OFA workshop.  I'd thought that a msgq (aka tag matcher class)
> object
> should be instantiated via a method of the fabric class.
> 
> red lines in the diagram indicated the pointee can be associated with the
> class
> being pointed to by the arrow, using the bind method of the class being
> pointed
> to.
> 
> the search_by_addr method of the msgq is for use with FID_RDM endpoints,
> while search_by_ep method is when the msgq is associated with multipled
> FID_MSG type endpoints.
> 
> Note the slide is a little old since the EC class has been divided now into
> a EQ and counter type completion notification mechanisms.
> 
> Hoping this will maybe help a little here.
> 
> Howard
> 
> 
> 
> On Thu, May 22, 2014 at 11:59 AM, Richard Graham <richardg at mellanox.com>
> wrote:
> 
> 
> 	Please see inline
> 
> 	-----Original Message-----
> 	From: Hefty, Sean [mailto:sean.hefty at intel.com]
> 	Sent: Thursday, May 22, 2014 12:43 PM
> 	To: Richard Graham; ofiwg at lists.openfabrics.org; ofiwg-
> mpi at lists.openfabrics.org
> 	Cc: Paul Grun (grun at cray.com); Liran Liss
> 	Subject: RE: Call today
> 
> 	With permission, copying mailing list on side thread that popped up.
> 
> 	I understand MPI has wild card receives.  But tagged semantics are
> useful even when associated with a generic endpoint concept, or a specific
> address.  Note the proposed endpoint concept is not necessarily bound to a
> specific piece of hardware, though it may be based on the provider
> implementation.  The tagged operations themselves may be implemented by
> hardware and are not restricted to being purely a software construct.
> 	[rich]  If the attempt here is to provide a building block that will
> map to different use-case scenarios, then need to have an architecture that
> will map well onto the areas of interest.  MPI is just one such upper level
> service, one that has been called out specifically in the context of the
> proposal you have been presenting.  So, following on this (the precise
> definition of end point is still rather fuzzy at this stage) in general,
> there is no such one-to-one mapping of and endpoint to an MPI matching
> context, but there can be an association of a matching context with one or
> more endpoints.  What I am suggesting here is that we keep data notions
> around data transfer orthogonal to what is done with the data (tag
> matching, in this case).  How the functionality is implemented  (hardware
> or not) is separate from how the stack in architected
> 
> 	Tagged interfaces, as well as other interfaces such as message
> queues, may still exist above the endpoint.  But that layering of
> interfaces seems better suited above the fabric interfaces (e.g. MPI),
> rather than included with it.  This seems more debatable to me though, and
> we could examine whether a domain or fabric object should have send/receive
> capabilities.
> 	[rich] Need to keep separate how data is transferred (perhaps with
> functions that we may call send/recv) from the ULP's use of this data
> (perhaps also using the a similar naming scheme of send/recv).
> 
> 	- Sean
> 
> 	> -----Original Message-----
> 	> From: Richard Graham [mailto:richardg at mellanox.com]
> 	> Sent: Wednesday, May 21, 2014 11:09 AM
> 	> To: Hefty, Sean
> 	> Cc: Paul Grun (grun at cray.com); Liran Liss
> 	> Subject: RE: Call today
> 	>
> 	> Tag matching as it comes to MPI semantics is not local to a given
> pair
> 	> of processes, e.g. MPI has a wild card receive that can take data
> from
> 	> any source, and therefore the matching context is broader than just
> a
> 	> single pair of source and destination.
> 	>
> 	> Rich
> 	>
> 	> -----Original Message-----
> 	> From: Hefty, Sean [mailto:sean.hefty at intel.com]
> 	> Sent: Wednesday, May 21, 2014 1:13 PM
> 	> To: Richard Graham
> 	> Cc: Paul Grun (grun at cray.com); Liran Liss
> 	> Subject: RE: Call today
> 	>
> 	> Tag matching, RMA, atomics, and message operations are currently
> 	> associated with an endpoint, but the functions are independent of
> the
> 	> communication protocol in use.  Conceptually, it seems reasonable
> to
> 	> think of tag matching as a merging of message and RMA write
> operations.
> 	>
> 	> I agree that an endpoint is associated with the data source/sink.
> 	> There is no implied mapping between a process and an endpoint.
> 	>
> 	>
> 	> > -----Original Message-----
> 	> > From: Richard Graham [mailto:richardg at mellanox.com]
> 	> > Sent: Tuesday, May 20, 2014 9:22 PM
> 	> > To: Hefty, Sean
> 	> > Cc: Paul Grun (grun at cray.com); Liran Liss
> 	> > Subject: RE: Call today
> 	> >
> 	> > I suppose that you could consider tag-matching as part of
> transport.
> 	> > However, I would argue that such protocols should be independent
> of
> 	> > whether or not a reliable or unreliable communication protocol is
> 	> > used
> 	> (at least
> 	> > when it comes to the tag support needed for MPI).    Also, I
> associate an
> 	> > end-point with either the source and/or the sync of data.  In MPI
> 	> > tag matching is associated with mpi-level (process,communicator)
> 	> > pair, and therefore the tag-matching context may be associated
> with
> 	> > many end-
> 	> points.
> 	> > I would therefore keep tag-matching as a separate concept.
> 	> >
> 	> > Rich
> 	> >
> 	> > -----Original Message-----
> 	> > From: Hefty, Sean [mailto:sean.hefty at intel.com]
> 	> > Sent: Tuesday, May 20, 2014 1:26 PM
> 	> > To: Richard Graham
> 	> > Cc: Paul Grun (grun at cray.com); Liran Liss
> 	> > Subject: RE: Call today
> 	> >
> 	> > Tag-matching is a transport object (protocol), so I do think it
> 	> > makes sense being associated with a transport level object (i.e.
> endpoint).
> 	> >
> 	> > I thought you were referring to the SRQ, which may or may not be
> a
> 	> > transport level object.  If the sharing of data buffer(s) among
> 	> > multiple connections is not considered a transport object, then I
> 	> > agree, it may make sense to have it be a separate object with its
> 	> > own
> 	> interfaces.
> 	> > Alternatively, it could also be a property of endpoints to share
> 	> > receive buffers.
> 	> >
> 	> > When the SRQ appears in the transport object (protocol), it may
> get
> 	> > more complex.
> 	> >
> 	> > For initial thoughts, sharing receive buffers could be handled
> by:
> 	> >
> 	> > 1. Creating an explicit SRQ object as a 'peer' to an endpoint.
> SRQ
> 	> > would have the ability to associate receive buffers with it.
> 	> > Endpoints would need to be associated with an SRQ to make use of
> it.
> 	> > 2. Create an SRQ 'endpoint' object.  A send-receive endpoint
> could
> 	> > be created from and inherent the SRQ interfaces.
> 	> > 3. Add an endpoint property to allow sharing data buffers.
> Shared
> 	> > buffers could be posted to a domain object, or, alternatively,
> any
> 	> endpoint.
> 	> >
> 	> > Ultimately, the question becomes a matter of where the 'post
> receive
> 	> > buffer' operation resides, and the behavior of any 'post receive
> buffer'
> 	> > call which may reside elsewhere.  E.g. SRQ::PostRecv() versus
> 	> > EP::PostRecv(), what is the behavior of EP::PostRecv() if buffer
> 	> > sharing is enabled?
> 	> >
> 	> > These assume SRQ as a non-transport object, or at least one that
> is
> 	> > not visible to the application.
> 	> >
> 	> >
> 	> >
> 	> > > Liran mentioned that you wanted me to repeat what I said - my
> only
> 	> > > comment was that we not couple transport (connection based
> 	> > > transport) with tag- matching (or any other object supported by
> 	> > > the
> 	> library).
> 	> > > These are two different concepts, and should be kept separate.
> 	> > >
> 	> > >
> 	> > >
> 	> > > Rich
> 
> 	_______________________________________________
> 	ofiwg mailing list
> 	ofiwg at lists.openfabrics.org
> 	http://lists.openfabrics.org/mailman/listinfo/ofiwg
> 
> 



More information about the ofiwg-mpi mailing list