These selections seem harmful in my measurements and don't seem to be
motivated by previous measurement data.
Signed-off-by: Joseph Schuchart <schuchart@icl.utk.edu>
Bcast: scatter_allgather and scatter_allgather_ring expect N_elem >= N_procs
Allreduce: rabenseifner expects N_elem >= pow2 nearest to N_procs
In all cases, the implementations will fall back to a linear implementation,
which will most likely yield the worst performance (noted for 4B bcast on 128 ranks)
Signed-off-by: Joseph Schuchart <schuchart@icl.utk.edu>
The mca parameters coll_tuned_*_algorithm are ignored unless coll_tuned_use_dynamic_rules is true so mention that in the description.
Signed-off-by: Joseph Schuchart <schuchart@icl.utk.edu>
Among many other things:
- Fix an imbalance bug in MPI_allgather
- Accept more human readable configuration files. We can now specify
the collective by name instead of a magic number, and the component
we want to use also by name.
- Add the capability to have optional arguments in the collective
communication configuration file. Right now the capability exists
for segment lengths, but is yet to be connected with the algorithms.
- Redo the initialization of all HAN collectives.
Cleanup the fallback collective support.
- In case the module is unable to deliver the expected result, it will fallback
executing the collective operation on another collective component. This change
make the support for this fallback simpler to use.
- Implement a fallback allowing a HAN module to remove itself as
potential active collective module, and instead fallback to the
next module in line.
- Completely disable the HAN modules on error. From the moment an error is
encountered they remove themselves from the communicator, and in case some
other modules calls them simply behave as a pass-through.
Communicator: provide ompi_comm_split_with_info to split and provide info at the same time
Add ompi_comm_coll_preference info key to control collective component selection
COLL HAN: use info keys instead of component-level variable to communicate topology level between abstraction layers
- The info value is a comma-separated list of entries, which are chosen with
decreasing priorities. This overrides the priority of the component,
unless the component has disqualified itself.
An entry prefixed with ^ starts the ignore-list. Any entry following this
character will be ingnored during the collective component selection for the
communicator.
Example: "sm,libnbc,^han,adapt" gives sm the highest preference, followed
by libnbc. The components han and adapt are ignored in the selection process.
- Allocate a temporary buffer for all lower-level leaders (length 2 segments)
- Fix the handling of MPI_IN_PLACE for gather and scatter.
COLL HAN: Fix topology handling
- HAN should not rely on node names to determine the ordering of ranks.
Instead, use the node leaders as identifiers and short-cut if the
node-leaders agree that ranks are consecutive. Also, error out if
the rank distribution is imbalanced for now.
Signed-off-by: Xi Luo <xluo12@vols.utk.edu>
Signed-off-by: Joseph Schuchart <schuchart@icl.utk.edu>
Signed-off-by: George Bosilca <bosilca@icl.utk.edu>
* first import of Bull specific modifications to HAN
* Cleaning, renaming and compilation fixing Changed all future into han.
* Import BULL specific modifications in coll/tuned and coll/base
* Fixed compilation issues in Han
* Changed han_output to directly point to coll framework output.
* The verbosity MCA parameter was removed as a duplicated of coll verbosity
* Add fallback in han reduce when op cannot commute and ppn are imbalanced
* Added fallback wfor han bcast when nodes do not have the same number of process
* Add fallback in han scatter when ppn are imbalanced
+ fixed missing scatter_fn pointer in the module interface
Signed-off-by: Brelle Emmanuel <emmanuel.brelle@atos.net>
Co-authored-by: a700850 <pierre.lemarinier@atos.net>
Co-authored-by: germainf <florent.germain@atos.net>
Reduce scatter block and reduce scatter algorithms were hitting
correctness issues for non commutative strided tests. We will revert to
the original default algorithms for those two collectives (basic linear
and non overlapping respectively) in the non commutative op case.
See #8010
Signed-off-by: William Zhang <wilzhang@amazon.com>
The default algorithm selections were out of date and not performing
well. After gathering data from OMPI developers, new default algorithm
decisions were selected for:
allgather
allgatherv
allreduce
alltoall
alltoallv
barrier
bcast
gather
reduce
reduce_scatter_block
reduce_scatter
scatter
These results were gathered using the ompi-collectives-tuning package
and then averaged amongst the results gathered from multiple OMPI
developers on their clusters.
You can access the graphs and averaged data here:
https://drive.google.com/drive/folders/1MV5E9gN-5tootoWoh62aoXmN0jiWiqh3
Signed-off-by: William Zhang <wilzhang@amazon.com>
The gather and scatter operations did not use the correct message size
(Only did datatype size * com size). This did not correctly reflect the
total message size and prevents fine tuning within a com size. This
patch multiplies the value by the number of elements sent.
Signed-off-by: William Zhang <wilzhang@amazon.com>
Use linear with sync alltoall algorithm for certain message/comm size
ranges. Does not affect default fixed decision, unless HPCX (with its
custom parameters) is used or corresponding mca is set.
Signed-off-by: Mikhail Brinskii <mikhailb@mellanox.com>
PR #5450 addresses MPI_IN_PLACE processing for basic collective algorithms.
But in conjunction with that, we need to check for MPI_IN_PLACE in tuned paths
as well before calling ompi_datatype_type_size() as otherwise we segfault.
MPI spec also stipulates to ignore sendcount and sendtype for Alltoall and
Allgatherv operations. So, extending the check to these algorithms as well.
Signed-off-by: Aravind Gopalakrishnan <Aravind.Gopalakrishnan@intel.com>
Implements butterfly algorithm for MPI_Reduce_scatter.
The algorithm can be used both by commutative and non-commutative operations, for power-of-two and non-power-of-two number of processes.
Signed-off-by: Mikhail Kurnosov <mkurnosov@gmail.com>
Implements butterfly algorithm for MPI_Reduce_scatter_block.
The algorithm can be used both by commutative and non-commutative
operations, for power-of-two and non-power-of-two number of processes.
Signed-off-by: Mikhail Kurnosov <mkurnosov@gmail.com>
- rename ompi_coll_base_reduce_scatter_block_basic to
more self descriptive ompi_coll_base_reduce_scatter_block_basic_linear
- fix the description of the coll_tuned_reduce_scatter_block_algorithm
MCA param
this fixes and documents previous open-mpi/ompi@0e8b35b615
MPI_Reduce_scatter_block used to be implemented by the coll/basic module only.
A new algo (recursive doubling) was recently introduced and can be used via the coll/tuned module,
but we never intended to make it the default algo.
In order to "restore" the previous default, the initial algo was moved from coll/basic to coll/base,
and is now used by default by coll/tuned.
Signed-off-by: Gilles Gouaillardet <gilles@rist.or.jp>
* Resolves#3705
* Components should link against the project level library to better
support `dlopen` with `RTLD_LOCAL`.
* Extend the `mca_FRAMEWORK_COMPONENT_la_LIBADD` in the `Makefile.am`
with the appropriate project level library:
```
MCA components in ompi/
$(top_builddir)/ompi/lib@OMPI_LIBMPI_NAME@.la
MCA components in orte/
$(top_builddir)/orte/lib@ORTE_LIB_PREFIX@open-rte.la
MCA components in opal/
$(top_builddir)/opal/lib@OPAL_LIB_PREFIX@open-pal.la
MCA components in oshmem/
$(top_builddir)/oshmem/liboshmem.la"
```
Note: The changes in this commit were automated by the script in
the commit that proceeds it with the `libadd_mca_comp_update.py`
script. Some components were not included in this change because
they are statically built only.
Signed-off-by: Joshua Hursey <jhursey@us.ibm.com>
in the base).
Correctly deal with persistent requests (they must be always freed when
they are stored in the request array associated with the communicator).
Always use MPI_STATUS_IGNORE for single request waiting functions.
