The direct modex operation is slow, especially at scale for even modestly-connected applications. Likewise, blocking in MPI_Init while we wait for a full modex to complete takes too long. However, as George pointed out, there is a middle ground here. We could kickoff the modex operation in the background, and then trap any modex_recv's until the modex completes and the data is delivered. For most non-benchmark apps, this may prove to be the best of the available options as they are likely to perform other (non-communicating) setup operations after MPI_Init, and so there is a reasonable chance that the modex will actually be done before the first modex_recv gets called.
Once we get instant-on-enabled hardware, this won't be necessary. Clearly, zero time will always out-perform the time spent doing a modex. However, this provides a decent compromise in the interim.
This PR changes the default settings of a few relevant params to make "background modex" the default behavior:
* pmix_base_async_modex -> defaults to true
* pmix_base_collect_data -> continues to default to true (no change)
* async_mpi_init - defaults to true. Note that the prior code attempted to base the default setting of this value on the setting of pmix_base_async_modex. Unfortunately, the pmix value isn't set prior to setting async_mpi_init, and so that attempt failed to accomplish anything.
The logic in MPI_Init is:
* if async_modex AND collect_data are set, AND we have a non-blocking fence available, then we execute the background modex operation
* if async_modex is set, but collect_data is false, then we simply skip the modex entirely - no fence is performed
* if async_modex is not set, then we block until the fence completes (regardless of collecting data or not)
* if we do NOT have a non-blocking fence (e.g., we are not using PMIx), then we always perform the full blocking modex operation.
* if we do perform the background modex, and the user requested the barrier be performed at the end of MPI_Init, then we check to see if the modex has completed when we reach that point. If it has, then we execute the barrier. However, if the modex has NOT completed, then we block until the modex does complete and skip the extra barrier. So we never perform two barriers in that case.
HTH
Ralph
Signed-off-by: Ralph Castain <rhc@open-mpi.org>
Adds:
- enabling/disabling of timings throught environment variable `OMPI_TIMING_ENABLE`
- output format: [file name]:[function name]:[description]: avg/min/max
- dynamically extending array of results for case then inited size was exhausted
- catch and collect errors
- cleanup
Note:
For use feature need to configure with `--enable-timings`
and set env `OMPI_TIMING_ENABLE = 1`
Signed-off-by: Boris Karasev <karasev.b@gmail.com>
This is an extension of OPAL timing framework that allows to use
MPI_reduce to provide the compact representation of the collected
timings throughout the whole application.
NOTE: the functionality is disabled now, it will be enabled after
the runtime verification.
Signed-off-by: Artem Polyakov <artpol84@gmail.com>
* Include a 'demo' component that shows some of the features.
* Currently has hooks for:
- MPI_Initialized
- top, bottom
- MPI_Init_thread
- top, bottom
- MPI_Finalized
- top, bottom
- MPI_Init
- top (pre-opal_init), top (post-opal_init), error, bottom
- MPI_Finalize
- top, bottom
* Other places in ompi can 'register' to hook into any one of these places
by passing back a component structure filled with function pointers.
* Add a `MCA_BASE_COMPONENT_FLAG_REQUIRED` flag to the MCA structure that
is checked by the `hook` framework. If a required, static component has
been excluded then the `hook` framework will fail to initialize.
- See note in `opal/mca/mca.h` as to why this is checked in the `hook`
framework and not in `opal/mca/base/mca_base_component_find.c`
Signed-off-by: Joshua Hursey <jhursey@us.ibm.com>
There are only five places in the non-daemon code paths where opal_hwloc_topology is currently referenced:
* shared memory BTLs (sm, smcuda). I have added a code path to those components that uses the location string
instead of the topology itself, if available, thus avoiding instantiating the topology
* openib BTL. This uses the distance matrix. At present, I haven't developed a method
for replacing that reference. Thus, this component will instantiate the topology
* usnic BTL. Uses the distance matrix.
* treematch TOPO component. Does some complex tree-based algorithm, so it will instantiate
the topology
* ess base functions. If a process is direct launched and not bound at launch, this
code attempts to bind it. Thus, procs in this scenario will instantiate the
topology
Note that instantiating the topology on complex chips such as KNL can consume
megabytes of memory.
Fix pernode binding policy
Properly handle the unbound case
Correct pointer usage
Do not free static error messages!
Signed-off-by: Ralph Castain <rhc@open-mpi.org>
Relax CPU usage pressure from the application processes when doing
modex and barrier in ompi_mpi_init.
We see significant latencies in SLURM/pmix plugin barrier progress
because app processes are aggressively call opal_progress pushing
away daemon process doing collective progress.
Newer versions of gcc have "poisoned" the __malloc_initialize_hook
name and it can no longer be used. Added a configure check and
protection around its usage.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
Add PMIx 2.0
Remove PMIx 1.1.4
Cleanup copying of component
Add missing file
Touchup a typo in the Makefile.am
Update the pmix ext114 component
Minor cleanups and resync to master
Update to latest PMIx 2.x
Update to the PMIx event notification branch latest changes
Update external as well
Revise the change: we still need the MPI_Barrier in MPI_Finalize when we use a blocking fence, but do use the "lazy" wait for completion. Replace the direct logic in MPI_Init with a cleaner macro
There is a potential race condition in MPI_Init() where an orte even
thread could be in a function that uses OPAL_THREAD_LOCK /
OPAL_THREAD_UNLOCK when ompi_mpi_init calls opal_set_using_threads().
Closesopen-mpi/ompi#1586
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit makes it possible to set relative priorities for
components. Before the addition of the patched component there was
only one component that would run on any system but that is no longer
the case. When determining which component to open each component's
query function is called and the one that returns the highest priority
is opened. The default priority of the patcher component is set
slightly higher than the old ptmalloc2/ummunotify component.
This commit fixes a long-standing break in the abstration of the
memory components. ompi_mpi_init.c was referencing the linux malloc
hook initilize function to ensure the hooks are initialized for
libmpi.so. The abstraction break has been fixed by adding a memory
base function that calls the open memory component's malloc hook init
function if it has one. The code is not yet complete but is intended
to support ptmalloc in 2.0.0. In that case the base function will
always call the ptmalloc hook init if exists.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit ensures the bml is always enabled whether or not it will
be used. This ensures that any available btls communicate their modex
so that they can be used for one-sided communication.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit rewrites both the mpool and rcache frameworks. Summary of
changes:
- Before this change a significant portion of the rcache
functionality lived in mpool components. This meant that it was
impossible to add a new memory pool to use with rdma networks
(ugni, openib, etc) without duplicating the functionality of an
existing mpool component. All the registration functionality has
been removed from the mpool and placed in the rcache framework.
- All registration cache mpools components (udreg, grdma, gpusm,
rgpusm) have been changed to rcache components. rcaches are
allocated and released in the same way mpool components were.
- It is now valid to pass NULL as the resources argument when
creating an rcache. At this time the gpusm and rgpusm components
support this. All other rcache components require non-NULL
resources.
- A new mpool component has been added: hugepage. This component
supports huge page allocations on linux.
- Memory pools are now allocated using "hints". Each mpool component
is queried with the hints and returns a priority. The current hints
supported are NULL (uses posix_memalign/malloc), page_size=x (huge
page mpool), and mpool=x.
- The sm mpool has been moved to common/sm. This reflects that the sm
mpool is specialized and not meant for any general
allocations. This mpool may be moved back into the mpool framework
if there is any objection.
- The opal_free_list_init arguments have been updated. The unused0
argument is not used to pass in the registration cache module. The
mpool registration flags are now rcache registration flags.
- All components have been updated to make use of the new framework
interfaces.
As this commit makes significant changes to both the mpool and rcache
frameworks both versions have been bumped to 3.0.0.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit removes the --with-mpi-thread-multiple option and forces
MPI_THREAD_MULTIPLE support. This cleans up an abstration violation
in opal where OMPI_ENABLE_THREAD_MULTIPLE determines whether the
opal_using_threads is meaningful. To reduce the performance hit on
MPI_THREAD_SINGLE programs an OPAL_UNLIKELY is used for the
check on opal_using_threads in OPAL_THREAD_* macros.
This commit does not clean up the arguments to the various functions
that take whether muti-threading support is enabled. That should be
done at a later time.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This fixes open-mpi/ompi@8b05f308f9
libmpi.so cannot be built (unresolved symbols) with configure'd with
--disable-mem-debug --disable-mem-profile --disable-memchecker --without-memory-manager
These changes fix issue https://github.com/open-mpi/ompi/issues/1336
- improve abstractions: opal/memory/linux component should be single place that opeartes with
Memory Allocation Hooks.
- avoid collisions in case dynamic component open/close: it is safe because it is linked statically.
- does not change original behaivour.
Update the configure logic for the new pmix120 component
ckpt
Get the pmix120 component to work - still not really registering or handling notifications, but infrastructure now operates
Cleanup some of the symbol scopes, and provide a more comprehensive rename.h file. Will pretty it up later - let's see how this works
Cleanup the rename files to use the pretty macros
to continue current default behavior.
Also add an MCA param pmix_base_collect_data to direct that the blocking fence shall return all data to each process. Obviously, this param has no effect if async_
modex is used.
Proposed extensions for Open MPI:
- If MPI_INITLIZED is invoked and MPI is only partially initialized,
wait until MPI is fully initialized before returning.
- If MPI_FINALIZED is invoked and MPI is only partially finalized,
wait until MPI is fully finalized before returning.
- If the ompi_mpix_allow_multi_init MCA param is true, allow MPI_INIT
and MPI_INIT_THREAD to be invoked multiple times without error (MPI
will be safely initialized only the first time it is invoked).
This commit adds support to the pml, mtl, and btl frameworks for
components to indicate at runtime that they do not support the new
dynamic add_procs behavior. At the high end the lack of dynamic
add_procs support is signalled by the pml using the new pml_flags
member to the pml module structure. If the
MCA_PML_BASE_FLAG_REQUIRE_WORLD flag is set MPI_Init will generate the
ompi_proc_t array passed to add_proc from ompi_proc_world () instead
of ompi_proc_get_allocated ().
Both cm and ob1 have been updated to detect if the underlying mtl and
btl components support dynamic add_procs.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit adds two new functions:
- ompi_proc_get_allocated - Returns all procs in the current job that
have already been allocated. This is used in init/finalize to
determine which procs to pass to add_procs/del_procs.
- ompi_proc_world_size - returns the number of processes in
MPI_COMM_WORLD. This may be removed in favor of callers just
looking at ompi_process_info.
The behavior of ompi_proc_world has been restored to return
ompi_proc_t's for all processes in the current job. The use of this
function is discouraged.
Code that was using ompi_proc_world() has been updated to make use of
the new functions to avoid the memory overhead of ompi_comm_world ().
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit adds an opal hash table to keep track of mapping between
process identifiers and ompi_proc_t's. This hash table is used by the
ompi_proc_by_name() function to lookup (in O(1) time) a given
process. This can be used by a BTL or other component to get a
ompi_proc_t when handling an incoming message from an as yet unknown
peer.
Additionally, this commit adds a new MCA variable to control the new
add_procs behavior: mpi_add_procs_cutoff. If the number of ranks in
the process falls below the threshold a ompi_proc_t is created for
every process. If the number of ranks is above the threshold then a
ompi_proc_t is only created for the local rank. The code needed to
generate additional ompi_proc_t's for a communicator is not yet
complete.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
Bring Slurm PMI-1 component online
Bring the s2 component online
Little cleanup - let the various PMIx modules set the process name during init, and then just raise it up to the ORTE level. Required as the different PMI environments all pass the jobid in different ways.
Bring the OMPI pubsub/pmi component online
Get comm_spawn working again
Ensure we always provide a cpuset, even if it is NULL
pmix/cray: adjust cray pmix component for pmix
Make changes so cray pmix can work within the integrated
ompi/pmix framework.
Bring singletons back online. Implement the comm_spawn operation using pmix - not tested yet
Cleanup comm_spawn - procs now starting, error in connect_accept
Complete integration
only define the unique fortran symbol depending on
- CAPS
- PLAIN
- SINGLE_UNDERSCORE
- DOUBLE_UNDERSCORE
and bind the f08 symbol to the uniquely defined C symbol.
Use real data structures to make the code simpler.
(perl script written by Jeff)
This commit fixes several vagrind errors. Included:
- installdirs did not correctly reinitialize all pointers to NULL
at close. This causes valgrind errors on a subsequent call to
opal_init_tool.
- several opal strings were leaked by opal_deregister_params which
was setting them to NULL instead of letting them be freed by the
MCA variable system.
- move opal_net_init to AFTER the variable system is initialized and
opal's MCA variables have been registered. opal_net_init uses a
variable registered by opal_register_params!
- do not leak ompi_mpi_main_thread when it is allocated by
MPI_T_init_thread.
- do not overwrite ompi_mpi_main_thread if it is already set (by
MPI_T_init_thread).
- mca_base_var: read_files was overwritting mca_base_var_file_list
even if it was non-NULL.
- mca_base_var: set all file global variables to initial states on
finalize.
- btl/vader: decrement enumerator reference count to ensure that it
is freed.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
We recognize that this means other users of OPAL will need to "wrap" the opal_process_name_t if they desire to abstract it in some fashion. This is regrettable, and we are looking at possible alternatives that might mitigate that requirement. Meantime, however, we have to put the needs of the OMPI community first, and are taking this step to restore hetero and SPARC support.
Properly setup the opal_process_info structure early in the initialization procedure. Define the local hostname right at the beginning of opal_init so all parts of opal can use it. Overlay that during orte_init as the user may choose to remove fqdn and strip prefixes during that time. Setup the job_session_dir and other such info immediately when it becomes available during orte_init.
Replace our old, clunky timing setup with a much nicer one that is only available if configured with --enable-timing. Add a tool for profiling clock differences between the nodes so you can get more precise timing measurements. I'll ask Artem to update the Github wiki with full instructions on how to use this setup.
This commit was SVN r32738.
