This commit updates the coll/basic component to correctly order sends
and receives for cartesian communicators with cyclic boundaries. This
addresses an issue identified by mpi-forum/mpi-issues#153. This issue
occurs when the size in any dimension is 1. This gives the same
neighbor in the positive and negative directions. The old code was
sending and receiving in the same order so the -1 buffer contained
the +1 result and vise-versa. The problem is addressed by using
unique tags for each send. This should cover both the case where
overtaking is allowed and is not allowed. The former case will be
possible is a MPI_Cart_create_with_info() call is added to the
standard.
Signed-off-by: Nathan Hjelm <hjelmn@google.com>
(cherry picked from commit 196a91e604)
open-mpi/ompi@0fe756d416 Introduced
a bug in coll/hcoll component. The ompi_requests allocated by
libhcoll would be treated as coll_base_nbc_request during
ompi_coll_base_retain_<> call. Afterwards this would lead to a
segv in the request cleanup.
Fix: since libhcoll interface does not distinguish between the
blocling/non-blocking requests use coll_base_nbc_request all the
time and initialize it properly in
coll/hcoll/get_coll_handle(). It is still within 2 cache lines.
Signed-off-by: Valentin Petrov <valentinp@mellanox.com>
a non blocking collective might return ompi_request_null, so we should not
retain anything in that case.
Signed-off-by: Gilles Gouaillardet <gilles@rist.or.jp>
(cherry picked from commit open-mpi/ompi@63d3ccde9d)
Since ompi_coll_base_nbc_request_t is to be used in an
opal_free_list_t, it must be returned into a "clean" state.
So cleanup some data in the callback completion subroutines.
This fixes a regression introduced in open-mpi/ompi@0fe756d416
Signed-off-by: Gilles Gouaillardet <gilles@rist.or.jp>
(cherry picked from commit open-mpi/ompi@0862c409f1)
base ompi_coll_libnbc_request_t on top of ompi_coll_base_nbc_request_t
to correctly support the retention of datatypes/operators
This fixes a regression introduced in open-mpi/ompi@0fe756d416
Signed-off-by: Gilles Gouaillardet <gilles@rist.or.jp>
(cherry picked from commit open-mpi/ompi@f8eef0fde9)
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>
(cherry picked from commit 404c480068)
MPI standard states a user MPI_Op and/or user MPI_Datatype can be free'd
after a call to a non blocking collective and before the non-blocking
collective completes.
Retain user (only) MPI_Op and MPI_Datatype when the non blocking call is
invoked, and set a request callback so they are free'd when the MPI_Request
completes.
Thanks Thomas Ponweiser for reporting this
Fixesopen-mpi/ompi#2151Fixesopen-mpi/ompi#1304
Signed-off-by: Gilles Gouaillardet <gilles@rist.or.jp>
(cherry picked from commit open-mpi/ompi@0fe756d416)
Signed-off-by: Aurelien Bouteiller <bouteill@icl.utk.edu>
Correctly bubble up errors in NBC collective operations
Signed-off-by: Aurelien Bouteiller <bouteill@icl.utk.edu>
The error field of requests needs to be rearmed at start, not at create
Signed-off-by: Aurelien Bouteiller <bouteill@icl.utk.edu>
(cherry picked from commit open-mpi/ompi@65660e5999)
If user sets HCOLL_EXTERNAL_UCM_EVENTS=1 then we try init opal
memory framework and register a mem release cb. Otherwise, rely on ucx.
Signed-off-by: Valentin Petrov <valentinp@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>
(cherry picked from commit 88d781056f)
The parameter passed to NBC_Return_handle() was incorrectly casted
and not dereferenced.
Thanks Yossi for the bug report.
Signed-off-by: Gilles Gouaillardet <gilles@rist.or.jp>
(cherry picked from commit open-mpi/ompi@8b51862fb2)
In the cleanup phase, it is possible for PtlMEUnlink() to return
PTL_IN_USE if the NIC is not done with the ME. This should not
be considered an error. This commit adds a retry loop around
PtlMEUnlink().
In some cases, the return value of PtlMEUnlink() and PtlCTFree()
was not checked at all. Check them with the same retry loop as
above.
Signed-off-by: Todd Kordenbrock <thkgcode@gmail.com>
(cherry picked from commit f3f2a826b4)
The call of MPI_Allgather with sendbuf and sendtype parameters equal to MPI_IN_PLACE and NULL correspondingly, produces the segmentation fault.
The problem is that sendtype is used even when sendbuf value is MPI_IN_PLACE. But according to the standard, sendtype and sendcount parameters should be ignored in this case.
Signed-off-by: Mikhail Kurnosov <mkurnosov@gmail.com>
(cherry picked from commit 540c2d1)
instead of invoking ompi_request_test_all(), that will end up
calling opal_progress() recursively, manually check the status
of the requests.
the same method is used in ompi_comm_request_progress()
Refs open-mpi/ompi#3901
Signed-off-by: Gilles Gouaillardet <gilles@rist.or.jp>
Use internal pack/unpack subroutines that operate on MPI_Aint instead of int
and hence solve some integer overflows.
Thanks Clyde Stanfield for reporting this issue.
Refs open-mpi/ompi#5383
Signed-off-by: Gilles Gouaillardet <gilles@rist.or.jp>
Current implementation of `coll/base/MPI_Scatter` is based on in-order binomial tree. This tree is right skewed and it provides good performance for a MPI_Gather operation. But for a MPI_Scatter operation left skewed binomial tree is effective.
Signed-off-by: Mikhail Kurnosov <mkurnosov@gmail.com>
The call of MPI_Gather with sendbuf and sendtype parameters equal to MPI_IN_PLACE and NULL correspondingly, produces the segmentation fault in the root process.
The problem is that sendtype is used even when sendbuf value is MPI_IN_PLACE. But according to the standard (page 150, line 37), sendtype and sendcount parameters should be ignored in this case.
Signed-off-by: Mikhail Kurnosov <mkurnosov@gmail.com>
The `nbc_i*` functions don't start communication, but create a request.
`nbc_*_init` are appropriate names for them.
Signed-off-by: KAWASHIMA Takahiro <t-kawashima@jp.fujitsu.com>
Persistent operation for `NBC_A2A_DISS` is not supported currently.
Though the algorithm is not selected at all currently, I put an
assertion not to select it by mistake.
Signed-off-by: KAWASHIMA Takahiro <t-kawashima@jp.fujitsu.com>
`NBC_Copy` shoud not be called in `MPI_*_INIT`.
`NBC_Sched_copy` should be called instead.
Signed-off-by: KAWASHIMA Takahiro <t-kawashima@jp.fujitsu.com>
Because a persistent reuqest does not free its `schedule` object
when the communication completes, the `NBC_Progress` function cannot
determine the completion using `schedule`.
Without this change, a hang occurs when the `NBC_Progress` function
is called recursively through the `NBC_Start_round` function.
Signed-off-by: KAWASHIMA Takahiro <t-kawashima@jp.fujitsu.com>
Now libnbc COLL supports persistent collectives and all `*_init`
functions of the COLL interface are available. So let's enable the
check of availability of those functions on a communicator creation.
Signed-off-by: KAWASHIMA Takahiro <t-kawashima@jp.fujitsu.com>
prepare the upcoming persistent collectives by pre-factoring some code
Signed-off-by: Gilles Gouaillardet <gilles@rist.or.jp>
fixup 808c3c62cd9475edd91ecde9d2d53b12e28b2c04
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>
Per MPI-3.1:8.7.1 p361:11-13, it's valid for MPI_FINALIZED to be
invoked during an attribute destruction callback (e.g., during the
destruction of keyvals on MPI_COMM_SELF during the very beginning of
MPI_FINALIZE). In such cases, MPI_FINALIZED must return "false".
Prior to this commit, we hung in FINALIZED if it were invoked during
a COMM_SELF attribute destruction callback in FINALIZE. See
https://github.com/open-mpi/ompi/issues/5084.
This commit converts the MPI_INITIALIZED / MPI_FINALIZED
infrastructure to use a single enum (ompi_mpi_state, set atomically)
to represent the state of MPI:
- not initialized
- init started
- init completed
- finalize started
- finalize past COMM_SELF destruction
- finalize completed
The "finalize past COMM_SELF destruction" state is what allows us to
return "false" from MPI_FINALIZED before COMM_SELF has been fully
destroyed / all attribute callbacks have been invoked.
Since this state is checked at nearly every MPI API call (to see if
we're outside of the INIT/FINALIZE epoch), care was taken to use
atomics to *set* the ompi_mpi_state value in ompi_mpi_init() and
ompi_mpi_finalize(), but performance-critical code paths can simply
read the variable without needing to use a slow call to an
opal_atomic_*() function.
Thanks to @AndrewGaspar for reporting the issue.
Signed-off-by: Jeff Squyres <jsquyres@cisco.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>
