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>
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>
To avoid fully initializing the osc/ucx component for MPI application
that are not using One-Sided functionality, the initialization happens
at the first MPI window creation.
This commit ensures atomicity of global state modifications.
Signed-off-by: Artem Polyakov <artpol84@gmail.com>
In common_ompi_aggregators calc_cost routine:
do not cast the real division to an int intermediately.
This patch removes the obsolete int variable c and assigns
the result of the P_a/P_x division directly to n_as.
With the intermediate int c variable, n_as gets 0 if P_a < P_x,
resulting in a division by 0 when computing n_s.
Signed-off-by: Harald Klimach <harald.klimach@uni-siegen.de>
this commmit fixes coverty warnings CID 1445198 and CID 1445197
For a reason that is a bit unclear to me, coverty only complained about the read
files, but the write operations had the same issue, so I fixed that within the
same commit as well.
Signed-off-by: Edgar Gabriel <egabriel@central.uh.edu>
external32 data representation is now support by ompio for everything
but non-blocking collective I/O operations. The support can further be improved
in a second step to limit the temporary buffer size (at least for blocking operations),
but it does work now for many scenarios.
Signed-off-by: Edgar Gabriel <egabriel@central.uh.edu>
introduce separate convertors for memory vs. file representation. Adjust the interfaces for decode_datatype to provide the convertor to be used for that.
Signed-off-by: Edgar Gabriel <egabriel@central.uh.edu>
the infrastructure put in place to manage cuda buffers is actually
a lot more generic than just for cuda buffers. Specifically, we ca
reuse much of the code to implement the external32 data representation.
This commit converts the code from common_ompio_cuda* to
common_ompio_buffer*. There are just very few places where we actually need to keep the OPAL_CUDA_SUPPORT ifdef in place.
Signed-off-by: Edgar Gabriel <egabriel@central.uh.edu>
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>
Atomic lock must progress local worker while obtaining the remote lock,
otherwise an active message which actually releases the lock might not
be processed while polling on local memory location.
Signed-off-by: Yossi Itigin <yosefe@mellanox.com>
The rdma_frag attached to the send request was not correctly released
upon request completion, leaking until MPI_Finalize. A quick solution
would have been to add RDMA_FRAG_RETURN at different locations on the
send request completion, but it would have unnecessarily made the
sendreq completion path more complex. Instead, I added the length to
the RDMA fragment so that it can be completed during the remote ack.
Be more explicit on the comment.
The rdma_frag can only be freed once when the peer forced a protocol
change (from RDMA GET to send/recv). Otherwise the fragment will be
returned once all data pertaining to it has been trasnferred.
Signed-off-by: George Bosilca <bosilca@icl.utk.edu>
The new routine transfers the data asynchronously from the source PE to all
PEs in the OpenSHMEM job. The routine returns immediately. The source and
target buffers are reusable only after the completion of the routine.
After the data is transferred to the target buffers, the counter object
is updated atomically. The counter object can be read either using atomic
operations such as shmem_atomic_fetch or can use point-to-point synchronization
routines such as shmem_wait_until and shmem_test.
Signed-off-by: Mikhail Brinskii <mikhailb@mellanox.com>
There are a couple MPI_Alltoallv calls in ad_gpfs_aggrs.c where the
send/recv data comes from places like req[r].lens, and the send
buffer and send displacements for example were being calculated as
sbuf = pick one of the reqs: req[bottom].lens
sdisps[r] = req[r].lens - req[bottom].lens
which might be okay if the .lens was data inside of req[] so they'd
all be close to each other. But each .lens field is just a pointer
that's malloced, so those addresses can be all over the place, so the
integer-sized sdisps[] isn't safe.
I changed it to have a new extra array sbuf and rbuf for those two
Alltoallv calls, and copied the data into the sbuf from the same
locations it used to be setting up the sdisps[] at, and after the
Alltoallv I copy the data out of the new rbuf into the same
locations it used to be setting up the rdisps[] at.
For what it's worth I was able to get this to fail -np 2 on a GPFS
filesystem with hints romio_cb_write enable. I didn't whittle the
test down to something small, but it was failing in an
MPI_File_write_all call.
Signed-off-by: Mark Allen <markalle@us.ibm.com>
abstract out the io_array structure to be used in common_ompio_build_io_array function.
This is preparation for a future component that would like to use the same function,
but not modify the io_array stored on the file handle itself.
Signed-off-by: Edgar Gabriel <egabriel@central.uh.edu>
In case of using a btl_put in ob1, the handle of the locally registered
memory is sent with a PUT control message. In the current master code
the sent handle is necessary the handle in the frag but if the handle
has been successfully registered in the request, the frag structure does
not have any valid handle and all fragments use the request one.
I suggest to check if the handle in the fragment is valid and if not to
send the handle from the request.
Signed-off-by: Brelle Emmanuel <emmanuel.brelle@atos.net>
In the case the btl_get fails Ob1 tries to fallback on btl_put first but
the return code was ignored. So the code fell back on both btl_put and
btl_send.
Signed-off-by: Brelle Emmanuel <emmanuel.brelle@atos.net>
This is not fixing any issue, it is simply preventing a sefault if the
communicator creation has not happened as expected. Thus, this code path
should never really be hit in a correct MPI application with a valid
communicator creation support.
Signed-off-by: George Bosilca <bosilca@icl.utk.edu>
- there was a set of UCX related issues reported which caused
by mmap API hooks conflicts. We added diagnostic of such
problems to simplify bug-resolving pipeline
Signed-off-by: Sergey Oblomov <sergeyo@mellanox.com>
