GitLab

We should always allocate target lock entry pool in win_init,
even though info no_locks is set to TRUE during window creation,
this is because that info can be set to FALSE by user after
the window creation.
Signed-off-by: Pavan Balaji <balaji@anl.gov>

Signed-off-by: Pavan Balaji <balaji@anl.gov>

Signed-off-by: Pavan Balaji <balaji@anl.gov>

Since it does not help on performance.
Signed-off-by: Pavan Balaji <balaji@anl.gov>

Originally we always dynamically allocate a request array
for the current RMA operation, since the current operation
might be streamed and needs multiple requests to track
each stream unit. However, in most cases where streaming is
not happening, we only needs one request for each operation
and does not need to dynamically allocate it. This patch
optimizes such case.
Signed-off-by: Pavan Balaji <balaji@anl.gov>

Signed-off-by: Pavan Balaji <balaji@anl.gov>

Originally we poke the progress engine at the end of RMA sync calls
if progress engine is never poked in this call before. The purpose
of this is to prevent possible deadlock problem. However, the deadlock
problem should only happen in self lock cases, if target is not myself,
it add unnecessary overhead to RMA sync calls. In this patch, we delete
those progress poking but only leave ones when target is myself.
Signed-off-by: Pavan Balaji <balaji@anl.gov>

In this patch, we add a reduce-scatter based algorithm in
MPI_Win_fence, which is triggered when number of processes
is at a small / medium value. When this algorithm is being
used, memory usage is O(P), but the ending FENCE only needs
to wait for local completion but does not need to wait for
remote completion. When number of processes is large, we
switch FENCE to the original barrier based algorithm, which
has O(1) memory usage, but needs to wait for the remote
completion in the ending FENCE.
Signed-off-by: Pavan Balaji <balaji@anl.gov>

Signed-off-by: Pavan Balaji <balaji@anl.gov>

Signed-off-by: Pavan Balaji <balaji@anl.gov>

On target side, after we receive the GACC/FOP packet, we should
first start sending back the data, then perform ACC computation.
By doing this issuing data and computation can be overlapped.
Signed-off-by: Pavan Balaji <balaji@anl.gov>

This optimization was missed in 7189bcde

.
Here we add this back so that when there is no iSSUED active win
or passive win, we ignore the while loop in RMA progress.
Signed-off-by: Pavan Balaji <balaji@anl.gov>

Originally the arguments passed to MPI_Win_create in this test
was wrong. This patch fixed this issue.
Signed-off-by: Pavan Balaji <balaji@anl.gov>

Signed-off-by: Halim Amer <aamer@anl.gov>

Signed-off-by: Halim Amer <aamer@anl.gov>

In the Nemesis implementation of Win_gather_info(), we allocate
a memory region on SHM to store window information for other
processes, so that all processes on the same node can share
those information. However, previously the memory size was
incorrectly set as O(node_comm_size), which should be O(comm_size).
This patch fixed this bug.
Signed-off-by: Min Si <msi@il.is.s.u-tokyo.ac.jp>
Signed-off-by: Pavan Balaji <balaji@anl.gov>

The original implementation in Win_flush_local counts number
of total local completion and remote completion needed to
wait, and then waiting for current local/remote completion
count to reach those values. There is a bug that we should
initialize the current count to zero in each while loop,
otherwise the targets that are already completed will be
count again and we failed to wait for some targets to be
completed. This patch fixes this issue.
Signed-off-by: Pavan Balaji <balaji@anl.gov>

No reviewer.
Signed-off-by: Pavan Balaji <balaji@anl.gov>

Signed-off-by: Junchao Zhang <jczhang@mcs.anl.gov>

No Reviewer

No reviewer

Lisandro Dalcin <dalcinl@gmail.com> reports that mpi4py's test suite
invokes MPIR_Add_finalize() 33 times.  It's been 6.5 years since we
doubled it, so bump it up once again.

Closes: #2272
Signed-off-by: Junchao Zhang <jczhang@mcs.anl.gov>

disable this memory-intensive test on 32 bit platforms

No Reviewer

resize and struct are the two type constructors that can set the LB and
UB markers on a type.  Struct, due to MPI-1 ideas, is a strange beast
(they adjust only if they are lower/higher than the old ones (!) )
but for resized it's clear that the markers shift.

Closes #2088
Signed-off-by: Pavan Balaji <balaji@anl.gov>

in deep types we might want to update the lb and ub, not simply
append/prepend two tuples to the flattened representation.
Signed-off-by: Pavan Balaji <balaji@anl.gov>

some libraries like HDF5 want to register their cleanup routines into
finalize.  these cleanup routines use MPI-IO, so they need to fire
before ROMIO cleans up.
Signed-off-by: Pavan Balaji <balaji@anl.gov>

Signed-off-by: Min Si <msi@il.is.s.u-tokyo.ac.jp>
Signed-off-by: Rob Latham <robl@mcs.anl.gov>

Signed-off-by: Sangmin Seo <sseo@anl.gov>

Signed-off-by: Sangmin Seo <sseo@anl.gov>

type promotions have resulted in a change to the device layer.

Ref: 1767
Signed-off-by: Pavan Balaji <balaji@anl.gov>
Signed-off-by: Sameh S Sharkawi <sssharka@us.ibm.com>

despite promoting types throughout the gather path, still had one case of
constructing structs with larger-than-int blocklens.  solution: borrow BigMPI
strategy and construct types-of-chunks to get around limitations.

Ref: #1767
Signed-off-by: Pavan Balaji <balaji@anl.gov>

The ongoing march towards 64-bit clean continues.  Address areas where large
product of two ints might have overflowed.

Ref: #1767
Signed-off-by: Pavan Balaji <balaji@anl.gov>

- preprocessor constants need parens
- which showed the "always fail" case wasn't big enough
- compiler warned about variables possibly being used uninitialized

Ref: #1767
Signed-off-by: Pavan Balaji <balaji@anl.gov>

Signed-off-by: Pavan Balaji <balaji@anl.gov>

Signed-off-by: Pavan Balaji <balaji@anl.gov>

Signed-off-by: Pavan Balaji <balaji@anl.gov>

No reviewer.

commit 83253a41

 triggerd a bunch of new warnings.  Take a different
approach.  For simplicity of implementation, do_accumulate_op is defined
as MPI_User_function.  We could split up internal routine and
user-provided routines, but that complicates the code for little
benefit:

Instead, keep do_accumlate_op with an int type, but check for overflow
before explicitly casting.  In many places the count is simply '1'.  In
stream processing there is an interal limit of 256k, so the assertion
should never fire.
Signed-off-by: Xin Zhao <xinzhao3@illinois.edu>

Encode request handles in unused tag bits, eliminating the need
for hash table.
Signed-off-by: Antonio Pena Monferrer <apenya@mcs.anl.gov>