June 2009 - mpiwg-tools - mailman3.cels.anl.gov

Re: [Mpi3-tools] MPI Forum/face-to-face meeting details?
by Martin Schulz 08 Jun '09

08 Jun '09

Hi all, Sorry for the late response - I was out of email contact last week and just got back online this morning. In the meantime, Rich has rescheduled the WG meetings to Wednesday morning - from 9am to 11:30am PDT. I hope that this earlier time slot will allow more people to participate. I asked for a phone line so that people can dial in, but I didn't get confirmation for this, yet. I will check on this once I get there and if this is available (which I hope) than I will send out another email with the dial-in information. Martin At 08:46 AM 6/8/2009, Marty Itzkowitz wrote: >Well? When is the tools face-to-face meeting? > > Thanks, > Marty > >On 6/1/09 8:47 AM, Jeff Squyres wrote: >>Martin -- >> >>Can you send out the meeting details for next week so that people >>know how/where to meet, and how to dial in? >> >>Thanks! >_______________________________________________ >Mpi3-tools mailing list >Mpi3-tools(a)lists.mpi-forum.org >http:// lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-tools _______________________________________________________________________ Martin Schulz, schulzm(a)llnl.gov, http://people.llnl.gov/schulz6 CASC @ Lawrence Livermore National Laboratory, Livermore, USA

1 0

Re: [Mpi3-tools] MPI Forum/face-to-face meeting details?
by Marty Itzkowitz 08 Jun '09

08 Jun '09

Well? When is the tools face-to-face meeting? Thanks, Marty On 6/1/09 8:47 AM, Jeff Squyres wrote: > Martin -- > > Can you send out the meeting details for next week so that people know > how/where to meet, and how to dial in? > > Thanks! >

1 0

Re: [Mpi3-tools] MPI handles debugging interface
by David Lecomber 03 Jun '09

03 Jun '09

Hi John and all, FWIW, all tools folks will do is just dlsym() lookups for the function, casting them to what we expect them to be, and calling them. So, only so long as the tool implementor has the same version of the header file as the library had, do we have type safety - hence extensions to existing functions and structs must be ruled out, extensions can only come via new functions and new structs. Can we be sure, in defining something strictly, that we've thought of everything? Ashley mentioned an extra byte count field, which is just one example we missed last year in defining an interface. The OMPI guys have a few extra fields at the end of a struct in the old MPI debug interface -- knowing that no-one else is using those. I guess what I had in mind was more a "Properties" style approach, you ask the debug interface what properties are supported, it gives you a list back, and you call a single function to get their (single) values - eg. communicator_get_attribute(foo). That'd give a debugger the ability to say "this communicator has attribute 'foo', I haven't a clue what it means, but I'm telling you it has value 'bar'" -- for any MPI added extension from day one. Overkill probably, but I just wanted to check opinions. Regards David On Wed, 2009-06-03 at 09:40 -0400, John DelSignore wrote: > IMHO, type safety is extremely important and any solution that is not type safe is a non-starter. > > Cheers, John D. > > > Jeff Squyres wrote: > > In chatting with David Lecomber while we're working on revising the MPI3 > > handle debugging interface per the recent discussions in this group > > (moving away from the One Big Struct approach), David asked a me a > > question that I thought might be appropriate for the larger group. > > > > Why use a bunch of individual query functions when we could have a > > single, much more general form that is more easily extensible? The > > query function would essentially be driven by an enum indicating the > > specific request, and the OUT arguments could be varargs. Perhaps > > something like this: > > > > ===== > > int mpidbg_comm_query_element(mpidbg_comm_handle_t handle, > > enum mpidbg_comm_query_element_t element, > > ...); > > ===== > > > > Here's a sample use case: > > > > ===== > > /* initial query to allow the DLL to bundle up *all* the data into a > > handle */ > > mpidbg_comm_query(image, image_info, ..., mpi_communicator, &handle); > > > > /* Query the "basic" communicator info from the handle */ > > mpidbg_comm_query_element(handle, MPIDBG_COMM_QUERY_BASIC, &flags, > > &rank, &size, ...etc.); > > > > /* Query the communicator process info from the handle */ > > mpidbg_comm_query_element(handle, MPIDBG_COMM_QUERY_PROCS, > > &num_local_procs, &local_procs, > > &num_remote_procs, &remote_procs); > > ===== > > > > +++ Advantage: this one function can be extended simply by adding new > > enum values, possibly in an MPI implementation-specific way (e.g., have > > some non-standard / value-add functions) > > > > --- Disadvantage: lose compiler type safety checking when > > building/debugging tools that use this interface. > > > > I don't have strong feelings either way. I think I'm marginally in > > favor of individual functions just because it gives type safety for > > tools authors. But I could be convinced to go the enum/varargs route if > > people feel strongly. > > > > Comments? > > > > --Jeff Squyres > > Cisco Systems > > > > _______________________________________________ > > Mpi3-tools mailing list > > Mpi3-tools(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-tools > > > > > _______________________________________________ > Mpi3-tools mailing list > Mpi3-tools(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-tools

1 0

Re: [Mpi3-tools] MPI handles debugging interface
by John DelSignore 03 Jun '09

03 Jun '09

IMHO, type safety is extremely important and any solution that is not type safe is a non-starter. Cheers, John D. Jeff Squyres wrote: > In chatting with David Lecomber while we're working on revising the MPI3 > handle debugging interface per the recent discussions in this group > (moving away from the One Big Struct approach), David asked a me a > question that I thought might be appropriate for the larger group. > > Why use a bunch of individual query functions when we could have a > single, much more general form that is more easily extensible? The > query function would essentially be driven by an enum indicating the > specific request, and the OUT arguments could be varargs. Perhaps > something like this: > > ===== > int mpidbg_comm_query_element(mpidbg_comm_handle_t handle, > enum mpidbg_comm_query_element_t element, > ...); > ===== > > Here's a sample use case: > > ===== > /* initial query to allow the DLL to bundle up *all* the data into a > handle */ > mpidbg_comm_query(image, image_info, ..., mpi_communicator, &handle); > > /* Query the "basic" communicator info from the handle */ > mpidbg_comm_query_element(handle, MPIDBG_COMM_QUERY_BASIC, &flags, > &rank, &size, ...etc.); > > /* Query the communicator process info from the handle */ > mpidbg_comm_query_element(handle, MPIDBG_COMM_QUERY_PROCS, > &num_local_procs, &local_procs, > &num_remote_procs, &remote_procs); > ===== > > +++ Advantage: this one function can be extended simply by adding new > enum values, possibly in an MPI implementation-specific way (e.g., have > some non-standard / value-add functions) > > --- Disadvantage: lose compiler type safety checking when > building/debugging tools that use this interface. > > I don't have strong feelings either way. I think I'm marginally in > favor of individual functions just because it gives type safety for > tools authors. But I could be convinced to go the enum/varargs route if > people feel strongly. > > Comments? > > --Jeff Squyres > Cisco Systems > > _______________________________________________ > Mpi3-tools mailing list > Mpi3-tools(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-tools > >

1 0

Re: [Mpi3-tools] MPI handles debugging interface
by Ashley Pittman 03 Jun '09

03 Jun '09

On Wed, 2009-06-03 at 08:59 -0400, Jeff Squyres wrote: > Why use a bunch of individual query functions when we could have a > single, much more general form that is more easily extensible? The > query function would essentially be driven by an enum indicating the > specific request, and the OUT arguments could be varargs. > +++ Advantage: this one function can be extended simply by adding new > enum values, possibly in an MPI implementation-specific way (e.g., > have some non-standard / value-add functions) The existing method can also be extended simply by adding new functions, possibly in a MPI implementation-specific way. > --- Disadvantage: lose compiler type safety checking when building/ > debugging tools that use this interface. A very valid point, it's far from easy to understand the interface currently and we need all the help we can get via type checking or any other means. > I don't have strong feelings either way. I think I'm marginally in > favor of individual functions just because it gives type safety for > tools authors. But I could be convinced to go the enum/varargs route > if people feel strongly. Personally I'm in favour of extensions being done with sympathy to the existing interface. Unless there is a significant reason to move, which I can't see, changing the communication mechanism is just going to add confusion and raise the barrier for people implementing either side of this interface. Ashley Pittman,

1 0

[Mpi3-tools] MPI handles debugging interface
by Jeff Squyres 03 Jun '09

03 Jun '09

In chatting with David Lecomber while we're working on revising the MPI3 handle debugging interface per the recent discussions in this group (moving away from the One Big Struct approach), David asked a me a question that I thought might be appropriate for the larger group. Why use a bunch of individual query functions when we could have a single, much more general form that is more easily extensible? The query function would essentially be driven by an enum indicating the specific request, and the OUT arguments could be varargs. Perhaps something like this: ===== int mpidbg_comm_query_element(mpidbg_comm_handle_t handle, enum mpidbg_comm_query_element_t element, ...); ===== Here's a sample use case: ===== /* initial query to allow the DLL to bundle up *all* the data into a handle */ mpidbg_comm_query(image, image_info, ..., mpi_communicator, &handle); /* Query the "basic" communicator info from the handle */ mpidbg_comm_query_element(handle, MPIDBG_COMM_QUERY_BASIC, &flags, &rank, &size, ...etc.); /* Query the communicator process info from the handle */ mpidbg_comm_query_element(handle, MPIDBG_COMM_QUERY_PROCS, &num_local_procs, &local_procs, &num_remote_procs, &remote_procs); ===== +++ Advantage: this one function can be extended simply by adding new enum values, possibly in an MPI implementation-specific way (e.g., have some non-standard / value-add functions) --- Disadvantage: lose compiler type safety checking when building/ debugging tools that use this interface. I don't have strong feelings either way. I think I'm marginally in favor of individual functions just because it gives type safety for tools authors. But I could be convinced to go the enum/varargs route if people feel strongly. Comments? -- Jeff Squyres Cisco Systems

1 0

[Mpi3-tools] Meeting minutes
by Pavan Balaji 01 Jun '09

01 Jun '09

Attendees: Chris Gottbrath (Totalview), Jeff Squyres (Cisco), John DelSignore (Totalview), Marty Itzkowitz (Sun) Minutes: We discussed the PMI initial draft to see if the concept itself was of interest to the working group. The main concern was that it was quite restrictive for different MPI implementations to follow a similar API. For example, Open MPI internally uses a binary representation which means that they have to convert to strings on each end to work with the PMI API, which might not be optimal. On the positive side, we discussed how PMI can be used as a tool to provide information about dynamic processes to other tools (e.g., the debugger) or as a mechanism to collect other system information (such as the topology, and other node/system information such as uptime/load). The general consensus was to have an API that is higher level than PMI that can be used by different tools to query the MPI implementation for information. This is slightly different from the PMI scope in that each process cannot add any information key for that process and any other process within MPI can read it. So, it is not clear if/how this can be done in a common way, but Pavan will talk to John @ Totalview to see if this is possible. -- Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

[Mpi3-tools] MPI Forum/face-to-face meeting details?
by Jeff Squyres 01 Jun '09

01 Jun '09

Martin -- Can you send out the meeting details for next week so that people know how/where to meet, and how to dial in? Thanks! -- Jeff Squyres Cisco Systems

1 0