August 2013 - mpiwg-tools - mailman3.cels.anl.gov

[Mpi3-tools] Updated MQD document
by Anh Vo 20 Aug '13

20 Aug '13

Hi all, This is the updated MQD document. Per our discussion a few weeks ago, Martin will present the changes at the September meeting before the 1st vote. The changes are shown with deleted text (strikethrough) and new text (red color). The figure had to be updated so I wasn't sure how to do that properly and thus I just put in the new figure. Perhaps we can have the old version open along side with the new version to show the changes in the figure? The changes to the figure are as follows: 1) The two example functions from within the tool's box had their names renamed to dbg_ instead of mqs_ to clarify that these are tool's internal funtions 2) The double-headed arrow with the caption "Interaction through callbacks" are changed to two single-headed arrows because the mechanism of invoking functions between the tool and the dll are different (one via callback and one via exported functions) I've checked in the changes to the depot --Anh

1 0

Re: [Mpi3-tools] matching pvar session and handles
by Kathryn Mohror 20 Aug '13

20 Aug '13

Hi Junchao, > To get a pvar handle, we pass in a session in MPI_T_PVAR_HANDLE_ALLOC(session, pvar_index, obj_handle, handle, count). > > It seems after allocation, a pvar handle should be bound to a session internally. > If so, why do we need the session argument in functions such as MPI_T_pvar_read(session, handle, buf)? Is it there to facilitate the implementation to make sure handle is from session? Off the top of my head, I can't remember what the rationale behind that design decision was. However, I imagine that passing in both arguments gives flexibility to the implementation in how it supports variables. For example, the handle might be an index into an array of variables valid for a particular session. In that case, the handle value wouldn't necessarily be unique across sessions. > If not, what happens if users pass in a mismatched "session, handle" pair? In this case, you get back a return code indicating that the handle is bad. This clarification will go into the errata: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/386 Kathryn > > Thanks > > --Junchao Zhang > _______________________________________________ > Mpi3-tools mailing list > Mpi3-tools(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-tools ______________________________________________________________ Kathryn Mohror, kathryn(a)llnl.gov, http://people.llnl.gov/mohror1 CASC @ Lawrence Livermore National Laboratory, Livermore, CA, USA

1 0

[Mpi3-tools] matching pvar session and handles
by Junchao Zhang 20 Aug '13

20 Aug '13

To get a pvar handle, we pass in a session in MPI_T_PVAR_HANDLE_ALLOC(session, pvar_index, obj_handle, handle, count). It seems after allocation, a pvar handle should be bound to a session internally. If so, why do we need the session argument in functions such as MPI_T_pvar_read(*session*, *handle*, buf)? Is it there to facilitate the implementation to make sure *handle* is from *session*? If not, what happens if users pass in a mismatched "session, handle" pair? Thanks --Junchao Zhang

1 0

Re: [Mpi3-tools] count in MPI_T_PVAR_HANDLE_ALLOC
by Kathryn Mohror 20 Aug '13

20 Aug '13

Hi Junchao, > At p578 of the standard, > > MPI_T_PVAR_HANDLE_ALLOC(session, pvar_index, obj_handle, handle, count) > ... > OUT count number of elements used to represent this variable (integer) > > It looks to me that pvar can have different classes, e.g., MPI_T_PVAR_CLASS_COUNTER, and datatypes, e.g., MPI_UNSIGNED. And a pvar can be an array (of length count) of such datatypes. > > I guess the common case is count=1. > > Is my understanding correct? Yes, your understanding is correct. Hope that helps, Kathryn > > Thanks > > --Junchao Zhang > _______________________________________________ > Mpi3-tools mailing list > Mpi3-tools(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-tools ______________________________________________________________ Kathryn Mohror, kathryn(a)llnl.gov, http://people.llnl.gov/mohror1 CASC @ Lawrence Livermore National Laboratory, Livermore, CA, USA

1 0

[Mpi3-tools] count in MPI_T_PVAR_HANDLE_ALLOC
by Junchao Zhang 20 Aug '13

20 Aug '13

Hello, At p578 of the standard, MPI_T_PVAR_HANDLE_ALLOC(session, pvar_index, obj_handle, handle, count) ... OUT count number of elements used to represent this variable (integer) It looks to me that pvar can have different classes, e.g., MPI_T_PVAR_CLASS_COUNTER, and datatypes, e.g., MPI_UNSIGNED. And a pvar can be an array (of length *count)* of such datatypes. I guess the common case is count=1. Is my understanding correct? Thanks --Junchao Zhang

1 0

Re: [Mpi3-tools] MPI_T session question
by Pavan Balaji 17 Aug '13

17 Aug '13

On 08/17/2013 02:11 PM, William Gropp wrote: > MPI, where it implements the performance variables, doesn't need to > store more than one - only in the session structure would there be > copies. I think you are saying the same thing I wrote in my email. > I view (without having participated) the design of the session as a > place to put all of the overhead out of the critical path, including > potential thread safety overheads. My comment about a high quality > implementation is that it is possible, with good and careful design, to > avoid most of the extra costs in the case of a single active session > (probably one indirection and a branch). Yes, of course. The single session case is easy. In fact, depending on how many variables we have, the multi-session case might not be too expensive either. I was just checking to make sure my understanding was correct. -- Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

Re: [Mpi3-tools] MPI_T session question
by William Gropp 17 Aug '13

17 Aug '13

MPI, where it implements the performance variables, doesn't need to store more than one - only in the session structure would there be copies. That's a big difference, because the performance variables may be on critical paths, and keeping their access simple and fast is vital. I view (without having participated) the design of the session as a place to put all of the overhead out of the critical path, including potential thread safety overheads. My comment about a high quality implementation is that it is possible, with good and careful design, to avoid most of the extra costs in the case of a single active session (probably one indirection and a branch). Bill William Gropp Director, Parallel Computing Institute Deputy Director for Research Institute for Advanced Computing Applications and Technologies Thomas M. Siebel Chair in Computer Science University of Illinois Urbana-Champaign On Aug 17, 2013, at 1:12 PM, Pavan Balaji wrote: > > Sure. That was my first thought as well, but that doesn't reduce the number of variables being stored in most cases. And will probably get messy in the THREAD_MULTIPLE case when we need to support reset of a variable on one session and read of the same value on another session. > > -- Pavan > > On 08/17/2013 10:14 AM, William Gropp wrote: >> I believe that the answer is yes, but note that in many/most/all cases, >> the session can record the value of the underlying MPI implementation >> variable at start time and then compute the difference. It does mean >> that the access in a session may not be as simple as a memory reference, >> though a high-quality implementation could optimize for a single session. >> >> Bill >> >> William Gropp >> Director, Parallel Computing Institute >> Deputy Director for Research >> Institute for Advanced Computing Applications and Technologies >> Thomas M. Siebel Chair in Computer Science >> University of Illinois Urbana-Champaign >> >> >> >> >> On Aug 17, 2013, at 10:08 AM, Pavan Balaji wrote: >> >>> Folks, >>> >>> If an application has multiple sessions and does a PVAR_START on all >>> sessions, is the intention that the MPI implementation will keep track >>> of each pvar on each of the sessions? Since when the pvar_start is >>> called for each session is different, the values of these variables >>> can be different. So is the MPI implementation expected to replicate >>> these variables per session? Is this true for the read-only ones as >>> well, since the start time can be different for each session? >>> >>> -- Pavan >>> >>> -- >>> Pavan Balaji >>> http://www.mcs.anl.gov/~balaji >>> _______________________________________________ >>> 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 >> > > -- > Pavan Balaji > http://www.mcs.anl.gov/~balaji

1 0

Re: [Mpi3-tools] MPI_T session question
by Pavan Balaji 17 Aug '13

17 Aug '13

Sure. That was my first thought as well, but that doesn't reduce the number of variables being stored in most cases. And will probably get messy in the THREAD_MULTIPLE case when we need to support reset of a variable on one session and read of the same value on another session. -- Pavan On 08/17/2013 10:14 AM, William Gropp wrote: > I believe that the answer is yes, but note that in many/most/all cases, > the session can record the value of the underlying MPI implementation > variable at start time and then compute the difference. It does mean > that the access in a session may not be as simple as a memory reference, > though a high-quality implementation could optimize for a single session. > > Bill > > William Gropp > Director, Parallel Computing Institute > Deputy Director for Research > Institute for Advanced Computing Applications and Technologies > Thomas M. Siebel Chair in Computer Science > University of Illinois Urbana-Champaign > > > > > On Aug 17, 2013, at 10:08 AM, Pavan Balaji wrote: > >> Folks, >> >> If an application has multiple sessions and does a PVAR_START on all >> sessions, is the intention that the MPI implementation will keep track >> of each pvar on each of the sessions? Since when the pvar_start is >> called for each session is different, the values of these variables >> can be different. So is the MPI implementation expected to replicate >> these variables per session? Is this true for the read-only ones as >> well, since the start time can be different for each session? >> >> -- Pavan >> >> -- >> Pavan Balaji >> http://www.mcs.anl.gov/~balaji >> _______________________________________________ >> 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 > -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0

Re: [Mpi3-tools] MPI_T session question
by William Gropp 17 Aug '13

17 Aug '13

I believe that the answer is yes, but note that in many/most/all cases, the session can record the value of the underlying MPI implementation variable at start time and then compute the difference. It does mean that the access in a session may not be as simple as a memory reference, though a high-quality implementation could optimize for a single session. Bill William Gropp Director, Parallel Computing Institute Deputy Director for Research Institute for Advanced Computing Applications and Technologies Thomas M. Siebel Chair in Computer Science University of Illinois Urbana-Champaign On Aug 17, 2013, at 10:08 AM, Pavan Balaji wrote: > Folks, > > If an application has multiple sessions and does a PVAR_START on all sessions, is the intention that the MPI implementation will keep track of each pvar on each of the sessions? Since when the pvar_start is called for each session is different, the values of these variables can be different. So is the MPI implementation expected to replicate these variables per session? Is this true for the read-only ones as well, since the start time can be different for each session? > > -- Pavan > > -- > Pavan Balaji > http://www.mcs.anl.gov/~balaji > _______________________________________________ > Mpi3-tools mailing list > Mpi3-tools(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-tools

1 0

[Mpi3-tools] MPI_T session question
by Pavan Balaji 17 Aug '13

17 Aug '13

Folks, If an application has multiple sessions and does a PVAR_START on all sessions, is the intention that the MPI implementation will keep track of each pvar on each of the sessions? Since when the pvar_start is called for each session is different, the values of these variables can be different. So is the MPI implementation expected to replicate these variables per session? Is this true for the read-only ones as well, since the start time can be different for each session? -- Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0