Skip to content
Jeff Squyres edited this page Sep 2, 2014 · 3 revisions

Notes from 5 Oct 2007

From Paris meeting and from lengthy phone conversation with Brian Barrett.

  • We specifically turned off epoll (and all other newer wait-on-a-bunch-of-fd's [WOABOF's] mechanisms) because they all have interesting / catastrophic failure modes when used with pty's on Solaris, OS X, and Linux ranging from segv's to kernel panics (!). Hence, since the IOF pipes in the orted to spawned processes may be pty's, we simply turned off all the advanced WOABOF's and fall back to select/poll. They appear to be turned off by a few simple #if 0's in libevent files; no configure updates should be required to turn them back on.

  • Brian thinks that we may be able to either use an existing libevent environment variable to force a run-time selection between select/poll and the WOABOF's or simply add our own internal API. This would let us have the orted use select/poll (even though many of its fd's will likely be real sockets [not pty's], the performance there is not very important, so it may be easier just to have all orted fd's use select/poll and have MPI app's use WOABOF's.

  • There are three modes to libevent's looping:

    • OPAL_EVLOOP_NONBLOCK: poll -- call select/poll/WOABOF with a zero timeout and check everything just once.
    • OPAL_EVLOOP_ONCE: block waiting for something to happen (e.g., signal is raised, any of the fd's are ready for their respective events, or the timeout expired).
    • OPAL_EVLOOP_ONELOOP: default mode used by the orted's (today) -- block for the default timeout (2sec?) waiting for something to happen.
  • George feels that Christian's proposal for revamping progress could be implemented in terms of libevent's functionality (i.e., no need to write anything/much new -- just use libevent properly). However, both George and Brian feel that there may be performance issues with our current libevent -- the signal/timer recomputations are expensive and it calls gettimeofday a few bazillion times.

    • Brian has been following the libevent mailing lists. He thinks that the new versions of libevent may have two things that interest us: 1) fixed/better signal/timer signal recomputations, and 2) call gettimeofday() a lot less.
    • Note that our libevent does use gettimeofday(), not a high resolution timer. We have an opal HR timer thingy (which is used in opal_progress()), but it is not used by libevent. Note that we do not have a solid mechanism for the intel chip high-resolution timer problem, namely: each CPU has its own indepdenent timer such that if a process bounces between processors, it may get wholly different/unrelated timer values (e.g., process A on processor 0 reads time T, but then process A bounces to processor 1, reads time again, and gets T-x -- whoops!). There are well-known algorithms to deal with this but we never bothered in OMPI because it didn't matter in the way we used the HR timers. Specifically, Brian says that there is some "funky" math in opal_progress() to deal with this problem, but all it does it guarantee that we don't timeout more than 2*desired_time_delta longer than we intend to (or something like that) - which is "good enough" for OMPI's present use of the HR timers (i.e., determining when to call libevent). However, if we move to Christian's proposal and start using libevent properly to regulate all events and timers (with calls to a cheap HR timer instead of gettimeofday), we will need to have the fix-the-intel-chip algorithms in place. Perhaps Christian can contribute here because he may have dealt with this in PSM already. Note that PPC and SPARC chips do not have this problem; they have HW-synchronized blocks so that it doesn't matter which processor you are on; a process will always get a synchronized view of the HR time.
    • In terms of merging in a new version of libevent to OMPI:
      • Note that the new libevent has infrastructure for non-blocking DNS and HTTP clients (not servers). We may or may not choose to import that stuff. Brian thinks that the libevent people will be splitting that stuff into a different library in the next release anyway.
      • A long, long time ago in a galaxy far, far away, someone (maybe even me...?) renamed all the libevent symbols and flags to have opal/OPAL prefixes. This makes importing a new version a real PITA. It may be far, far simpler to have a .h file with #define's that simply converts foo to opal_foo for all the symbols that matter (let's stop renaming flags that are not linkable symbols -- that really isn't necessary). This will allow us to bring in new versions of libevent with much fewer manual modifications. Brian started this work with event_rename.h, but really only did new symbols that were recently added, he didn't take a comprehensive effort to fix everything.
      • Libevent's threading model is quite similar to POSIX foo_r() functionality: you pass in a pointer to the data structure that may be modified and it's the caller's responsibility to ensure that thread safety w.r.t. data structures occurs properly. However, OMPI's threading model / use of libevent kinda violates this model. Specifically, opal_progress() has one data structure that it uses for libevent. But multiple threads may call opal_progress() similtaneously, so multiple threads may end up calling libevent with the same data structure simultaneously. Tim Woodall added fine-grained locking in the parts of libevent that we use to make this all work properly. HOWEVER, if we start using WOABOF's and/or we import new versions of libevent, we either need to change opal_progress()'s use of this data structure or we need to present/extend the locking scheme that Tim Woodall implemented.
      • We may also want to change the use of gettimeofday to a cheap high-resolution timer (preferably with a clever #define kind of way, if possible, to make future SVN merges of new versions of libevent less painful). See above for a discussion of the opal HR timer stuff.
  • Random opal_progress() note: the functions opal_progress_event_users_increment() and opal_progress_event_users_decrement() are used to count the number of "users" of libevent (e.g., if anyone needs to have opal_progress call libevent or not). If num_users>0, opal_progress will call libevent frequently. If num_users==0, opal_progress calls libevent infrequently. Hence, the openib BTL increments the num_users count during wireup and then decrements it when wireup is done (in the openib OOB connect pseudo-component).

  • Random yield/opal_progress/libevent note: orted takes all the defaults for yield/opal_progress/libevent. ompi_mpi_init() plays games -- it sets up yield()ing behavior and frequent libevent invocations so that we can cooporate with the orted and shared memory setup (which depend on multiple processes -- potentially more processes than the number of processors -- sharing CPU cycles nicely) and allow all processes on a single host to complete these tasks in a timely fashion. At the end of ompi_mpi_init(), we set the yield/libevent behavior to be whatever is actually desired for the MPI job.

  • Random opal_progress() note: yield() should only be called if:

    • "I want to yield" global behavior is enabled
    • Nothing else happened in progress (to include libevent, I think?)
  • Action item: Brian just recently wrote a progress handler for the orted for routing reasons; there is a race condition where a message may arrive in an orted for a process before there is a route setup to that process yet. Hence, we have to queue up the message and then check later to see if a route has been setup yet (see r16321). Brian implemented this as a simple progress function simply because he ran out of time to re-look up how to use the timer functions. However, this would conflict with the idea of setting the orted to EVLOOP_ONCE because we would also need to check this progress function (i.e., flipping between progress and libevent). Implementing it as a time would be better because then libevent can handle the whole thing -- no need to flip between progress and libevent. Brian recommends fixing this up to use a timed libevent in conjunction with the change to make the orted run in EVLOOP_ONCE (e.g., have it check for the new route ever .25, .5, or 1 sec or something).

Clone this wiki locally