This creates a really bad scaling behavior. Users have found a nearly 20% launch time differential between mpirun and PMI, with PMI being the slower method. Some of the problem is attributable to poor exchange algorithms in RM's like Slurm and Alps, but we make things worse by calling "get" so many times.
Nathan (with a tad advice from me) has attempted to alleviate this problem by reducing the number of "get" calls. This required the following changes:
* upon first request for data, have the OPAL db pmi component fetch and decode *all* the info from a given remote proc. It turned out we weren't caching the info, so we would continually request it and only decode the piece we needed for the immediate request. We now decode all the info and push it into the db hash component for local storage - and then all subsequent retrievals are fulfilled locally
* reduced the amount of data by eliminating the exchange of the OMPI_ARCH value if heterogeneity is not enabled. This was used solely as a check so we would error out if the system wasn't actually homogeneous, which was fine when we thought there was no cost in doing the check. Unfortunately, at large scale and with direct launch, there is a non-zero cost of making this test. We are open to finding a compromise (perhaps turning the test off if requested?), if people feel strongly about performing the test
* reduced the amount of RTE data being automatically fetched, and fetched the rest only upon request. In particular, we no longer immediately fetch the hostname (which is only used for error reporting), but instead get it when needed. Likewise for the RML uri as that info is only required for some (not all) environments. In addition, we no longer fetch the locality unless required, relying instead on the PMI clique info to tell us who is on our local node (if additional info is required, the fetch is performed when a modex_recv is issued).
Again, all this only impacts direct launch - all the info is provided when launched via mpirun as there is no added cost to getting it
Barring objections, we may move this (plus any required other pieces) to the 1.7 branch once it soaks for an appropriate time.
This commit was SVN r29040.
non-contiguous converter. We can't "convert on the fly" because the #
of bytes requested may not divide evenly into the convertor data type.
This commit was SVN r29014.
spawning a process using MPI_Comm_spawn. For this, the first operation has to
be collective which we can not guarantuee outside of the MPI_File_open
operation.
This commit was SVN r29008.
improvements:
* Fix minor memory leaks during component_init
* Ensure that an initialization loop does not underflow an unsigned int
* Improve mlock limit checking
* Fix set of BTL modules created during component_init when failing to
get QP resources or otherwise excluding some (but not all) usnic
verbs devices
* Fix/improve error messages to be consistent with other Cisco
documentation
* Randomize the initial sliding window sequence number so that we
silently drop incoming frames from previous jobs that still have
existant processes in the middle of dying (and are still
transmitting)
* Ensure we don't break out of add_procs too soon and create an
asymetrical view of what interfaces are available
This commit was SVN r28975.
mpi_leave_pinned_pipeline
This change should improve performance is the non-pinned case where
the same memory region is involved in multiple simultaneous transfers.
cmr=v1.7.3:reviewer=brbarret
This commit was SVN r28973.
Working on faster algorithms for tuned that will come at a later time.
cmr=v1.7.3:ticket=trac:2965
This commit was SVN r28952.
The following Trac tickets were found above:
Ticket 2965 --> https://svn.open-mpi.org/trac/ompi/ticket/2965
Use the new sysfs files to check that there are enough VFs, QPs, and
CQs for all the MPI processes on this server.
Move the checking code into its own subroutine to make it smaller and
easier to read/grok.
This commit was SVN r28937.
Brian (rightfully) hit me on the head with the
don't-use-ORTE-use-the-rte-framework clue bat; the usnic BTL now
nicely plays with the RTE framework.
This commit was SVN r28907.
Note that the PMI RTE still doesn't listen for asynchronous errors, so
the error handler still won't ever actually do anything :).
This commit was SVN r28886.
The following SVN revision numbers were found above:
r28852 --> open-mpi/ompi@e4e678e234
This BTL accesses the Cisco usNIC Linux device via the Linux verbs
API via Unreliable Datagram queue pairs. A few noteworthy points:
* This BTL does most of its own fragmentation; it tells the PML that
it has a very high max_send_size (much higher than the network
MTU).
* Since UD fragments are, by definition, unreliable, the usnic BTL
handles all of its own reliability via a sliding window approach
using the opal_hotel construct and many tricks stolen from the
corpus of knowledge surrounding efficient TCP.
* There is a fun PML latency-metric based optimization for NUMA
awareness of short messages.
* Note that this is ''not'' a generic UD verbs BTL; it is specific to
the Cisco usNIC device.
This commit was SVN r28879.
George and I were talking about ORTE's error handling the other day in regards to the right way to deal with errors in the updated OOB. Specifically, it seemed a bad idea for a library such as ORTE to be aborting the job on its own prerogative. If we lose a connection or cannot send a message, then we really should just report it upwards and let the application and/or upper layers decide what to do about it.
The current code base only allows a single error callback to exist, which seemed unduly limiting. So, based on the conversation, I've modified the errmgr interface to provide a mechanism for registering any number of error handlers (this replaces the current "set_fault_callback" API). When an error occurs, these handlers will be called in order until one responds that the error has been "resolved" - i.e., no further action is required - by returning OMPI_SUCCESS. The default MPI layer error handler is specified to go "last" and calls mpi_abort, so the current "abort" behavior is preserved unless other error handlers are registered.
In the register_callback function, I provide an "order" param so you can specify "this callback must come first" or "this callback must come last". Seemed to me that we will probably have different code areas registering callbacks, and one might require it go first (the default "abort" will always require it go last). So you can append and prepend, or go first. Note that only one registration can declare itself "first" or "last", and since the default "abort" callback automatically takes "last", that one isn't available. :-)
The errhandler callback function passes an opal_pointer_array of structs, each of which contains the name of the proc involved (which can be yourself for internal errors) and the error code. This is a change from the current fault callback which returned an opal_pointer_array of just process names. Rationale is that you might need to see the cause of the error to decide what action to take. I realize that isn't a requirement for remote procs, but remember that we will use the SAME interface to report RTE errors internal to the proc itself. In those cases, you really do need to see the error code. It is legal to pass a NULL for the pointer array (e.g., when reporting an internal failure without error code), so handlers must be prepared for that possibility. If people find that too burdensome, we can remove it.
Should we ever decide to create a separate callback path for internal errors vs remote process failures, or if we decide to do something different based on experience, then we can adjust this API.
This commit was SVN r28852.
- extend the framework API
- remove the dummy component, not require anymore
- add four components to perform the actual job.
This commit was SVN r28828.
soon:
- add a new abstraction layer to be used internally for some operations
- add a new mca parameter to control lazy intialization of shared file
pointer structures
This commit was SVN r28826.
