On Thu, Dec 17, 2015 at 4:32 AM, Marc-Andre Hermanns < [email protected]> wrote:

Hi Jeff,

at the moment we don't handle MPI_THREAD_MULTIPLE at all. But we want to get there ;-)

You should vote for endpoints, as this may help you out here, particularly if users start mapping endpoints 1:1 w/ threads.

Here is a short recollection of what we do/need. Sorry for the folks who know/read this already in other context:

We use, what we call "parallel replay" to analyze large event traces in parallel. Each thread has its own stream of events, such as enter and exit for tracking the calling context as well as send and receive for communication among ranks.

We analyze on the same scale as the measurement, thus we have one thread per thread-local trace. Each thread processes its own thread-local trace. When encountering a communication event, it re-enacts this communication using the recorded communication parameters (rank, tag, comm). A send event leads to an issued send, a receive event leads to an issued receive.

It is critical that during the analysis, the message matching is identical to the original application. However, we do not re-enact any computational time, that is the temporal distance between sends and receives is certainly different from the original application. As a consequence, while two sends may have some significant temporal distance in the original measurement, they could be issued right after each other during the analysis.

Markus Geimer and I believe that creating some sort of a sequence number during measurement could help matching the right messages during the analysis, as a process could detect that it got mismatched messages and communicate with other threads to get the correct one.

It is unclear, however, how to achieve this:

a) Sending an additional message within the MPI wrapper at measurement time may lead to invalid matchings, as the additional message may be received by a different thread.

b) Creating a derived datatype on the fly to add tool-level data to the original payload may induce a large overhead in practically _every_ send & receive operation and perturb the measurement.

You should evaluate this experimentally. I wrote a simple test ( https://github.com/jeffhammond/BigMPI/blob/master/test/perf/typepiggy.c) and measured 1.5 us per call of overhead to create a datatype. That is not significant except for very small messages. Jeff The best idea in the room so far is some callback mechanism into the

MPI implementation that generates matching information both on sender and receiver side to generate some form of a sequence number that can then be saved during measurement. If available on both sender and receiver this information could then be used to fix incorrect matching during the analysis.

Cheers, Marc-Andre

On 16.12.15 16:40, Jeff Hammond wrote:

...

How do you handle MPI_THREAD_MULTIPLE? Understanding what your tool does there is a good starting point for this discussion.

Jeff

On Wed, Dec 16, 2015 at 1:37 AM, Marc-Andre Hermanns <[email protected] <mailto:[email protected]>> wrote:

Hi all,

CC: Tools-WG, Markus Geimer (not on either list)

sorry for starting a new thread and being so verbose, but I subscribed just now. I quoted Dan, Jeff, and Jim from the archive as appropriate.

First, let me state that we do not want to prevent this assertion in any way. For us as tools provider it is just quite a brain tickler on how to support this in our tool and in general.

Dan wrote: >>> [...] The basic problem is that message matching would be >>> non-deterministic and it would be impossible for a tool to show >>> the user which receive operation satisfied which send operation >>> without internally using some sort of sequence number for each >>> send/receive operation. [...] >>> >>> My responses were: >>> 1) the user asked for this behaviour so the tool could simply >>> gracefully give up the linking function and just state the >>> information it knows > Giving up can only be a temporary solution for tools. The user wants to use this advanced feature, thus just saying: "Hey, what you're doing is too sophisticated for us. You are on your own now." is not a viable long-term strategy.

>>> 2) the tool could hook all send and receive operations and >>> piggy-back a sequence number into the message header

We discussed piggy-backing within the tools group some time in the past, but never came to a satisfying way of how to implement this. If, in the process of reviving the discussion on a piggy-backing interface, we come to a viable solution, it would certainly help with the our issues with message matching in general.

Scalasca's problem here is that we need to detect (and partly recreate) the exact order of message matching to have the correct message reach the right receivers.

>>> 3) the tool could hook all send and receive operations and >>> serialise them to prevent overtaking

This is not an option for me. A "performance tool" should strive to measure as close to the original behavior as possible. Changing communication semantics just to make a tool "work" would have too great of an impact on application behavior. After all, if it would have only little impact, why should the user choose this option in the first place.

Jeff wrote: >> Remember that one of the use cases of allow_overtaking is applications that >> have exact matching, in which case allow_overtaking is a way of turning off >> a feature that isn't used, in order to get a high-performing message queue >> implementation. In the exact matching case, tools will have no problem >> matching up sends and recvs.

This is true. If the tools can identify this scenario, it could be supported by current tools without significant change. However, as it is not generally forbidden to have inexact matching (right?), it is unclear on how the tools would detect this.

What about an additional info key a user can set in this respect:

exact_matching => true/false

in which the user can state whether it is indeed a scenario of exact matching or not. The tool could check this, and issue a warning.

>> If tools cannot handle MPI_THREAD_MULTIPLE already, then I don't really >> care if they can't support this assertion either.

Not handling MPI_THREAD_MULTIPLE generally is not carved in stone. ;-)

As I said, we (Markus and I) see this as a trigger to come to a viable solution for tools like ours to support either situation.

>> And in any case, such tools can just intercept the info operations and >> strip this key if they can't support it.

As I wrote above in reply to Dan, stripping options that influence behavior is not a good option. I, personally, would rather bail out than (silently) change messaging semantics. I can't say what Markus' take on this is.

Jim wrote: > I don't really see any necessary fix to the proposal. We could add an > advice to users to remind them that they should ensure tools are compatible > with the info keys. And the reverse advice to tools writers that they > should check info keys for compatibility.

I would second this idea, while emphasizing the burden to be on the tool to check for this info key (and potentially others) and warn the user of "undersupport".

Cheers, Marc-Andre -- Marc-Andre Hermanns Jülich Aachen Research Alliance, High Performance Computing (JARA-HPC) Jülich Supercomputing Centre (JSC)

Schinkelstrasse 2 52062 Aachen Germany

Phone: +49 2461 61 2509 | +49 241 80 24381 Fax: +49 2461 80 6 99753 www.jara.org/jara-hpc <http://www.jara.org/jara-hpc> email: [email protected] <mailto:[email protected]>

_______________________________________________ mpiwg-p2p mailing list [email protected] <mailto:[email protected]> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-p2p

-- Jeff Hammond [email protected] <mailto:[email protected]> http://jeffhammond.github.io/

_______________________________________________ mpiwg-p2p mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-p2p

-- Marc-Andre Hermanns Jülich Aachen Research Alliance, High Performance Computing (JARA-HPC) Jülich Supercomputing Centre (JSC)

Schinkelstrasse 2 52062 Aachen Germany

Phone: +49 2461 61 2509 | +49 241 80 24381 Fax: +49 2461 80 6 99753 www.jara.org/jara-hpc email: [email protected]

_______________________________________________ mpiwg-p2p mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-p2p

-- Jeff Hammond [email protected] http://jeffhammond.github.io/