mark the "self" peer OMPI_OSC_RDMA_PEER_LOCAL_BASE when
the window is dynamically created and use_cpu_atomics is set
in order to correctly handle communications to self.
Thanks Bart Janssens for reporting this issue.
Refs. open-mpi/ompi#6394
Signed-off-by: Gilles Gouaillardet <gilles@rist.or.jp>
The Open MPI code base assumed that asprintf always behaved like
the FreeBSD variant, where ptr is set to NULL on error. However,
the C standard (and Linux) only guarantee that the return code will
be -1 on error and leave ptr undefined. Rather than fix all the
usage in the code, we use opal_asprintf() wrapper instead, which
guarantees the BSD-like behavior of ptr always being set to NULL.
In addition to being correct, this will fix many, many warnings
in the Open MPI code base.
Signed-off-by: Brian Barrett <bbarrett@amazon.com>
The aggregation code in osc/rdma is currently broken and will likely
not be reused. This commit cleans it out.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit fixes a crash that occurs when using btl/vader as an RDMA
btl. This btl supports using CPU atomics and does not support using
the btl for self communication so we must use the local memory
optimizations in osc/rdma.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit adds a new MCA variable to set the location of the backing
store: osc_rdma_backing_directory. The default on Linux has been
changed to use /dev/shm to improve performance in cases where /tmp is
not a tmpfs.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit fixes a bug is osc/rdma that can occur if the total size
of the shared memory segment gets larger than 4 GiB. The bug was
caused by a typo. The type of my_base_offset should have been size_t
not int.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit is a large update to the osc/rdma component. Included in
this commit:
- Add support for using hardware atomics for fetch-and-op and single
count accumulate when using the accumulate lock. This will improve
the performance of these operations even when not setting the
single intrinsic info key.
- Rework how large accumulates are done. They now block on the get
operation to fix some bugs discovered by an IBM one-sided test. I
may roll back some of the changes if the underlying bug in the
original design is discovered. There appear to be no real
difference (on the hardware this was tested with) in performance so
its probably a non-issue. References #2530.
- Add support for an additional lock-all algorithm: on-demand. The
on-demand algorithm will attempt to acquire the peer lock when
starting an RMA operation. The lock algorithm default has not
changed. The algorithm can be selected by setting the
osc_rdma_locking_mode MCA variable. The valid values are two_level
and on_demand.
- Make use of the btl_flush function if available. This can improve
performance with some btls.
- When using btl_flush do not keep track of the number of put
operations. This reduces the number of atomic operations in the
critical path.
- Make the window buffers more friendly to multi-threaded
applications. This was done by dropping support for multiple
buffers per MPI window. I intend to re-add that support once the
underlying performance bug under the old buffering scheme is
fixed.
- Fix a bug in request completion in the accumulate, get, and put
paths. This also helps with #2530.
- General code cleanup and fixes.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit fixes the following bugs:
- Allow a btl to be used for communication if it can communicate with
all non-self peers and it supports global atomic visibility. In
this case CPU atomics can be used for self and the btl for any
other peer.
- It was possible to get into a state where different threads of an
MPI process could issue conflicting accumulate operations to a
remote peer. To eliminate this race we now update the peer flags
atomically.
- Queue up and re-issue put operations that failed during a BTL
callback. This can occur during an accumulate operation. This was
an unhandled error case.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
higher priority than rdma and default to psm2.
Context: the Intel Omni-path driver (hfi1) has verbs support, so the openib
btl is available to use. However, at a bad performance. Without this
change osc rdma using btl openib is the default choice when running on Intel
Omni-path, with a lower performance than osc pt2pt over mtl psm2.
Signed-off-by: Matias A Cabral <matias.a.cabral@intel.com>
The data endpoint was not being set correctly for local peers in some
cases. This commit fixes the bug and cleans the associated code to
simplify the logic.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
See bug report
https://github.com/open-mpi/ompi/issues/3548
If a 1sided test is launched -host hostA:2,hostB:1 some of the ranks
call allocate_state_single() and others call allocate_state_shared().
These functions were producing different values for module->state_size
but that's used when they lookup peer info from each other in
ompi_osc_rdma_peer_setup() so they need to all have matching
module->state_offset values.
This change adds a few unused bytes in the memory allocate_state_single()
creates so it matches.
Signed-off-by: Mark Allen <markalle@us.ibm.com>
The expected sequence of events for processing info during object creation
is that if there's an incoming info arg, it is opal_info_dup()ed into the obj
at obj->s_info first. Then interested components register callbacks for
keys they want to know about using opal_infosubscribe_infosubscribe().
Inside info_subscribe_subscribe() the specified callback() is called with
whatever matching k/v is in the object's info, or with the default. The
return string from the callback goes into the new k/v stored in info, and
the input k/v is saved as __IN_<key>/<val>. It's saved the same way
whether the input came from info or whether it was a default. A null return
from the callback indicates an ignored key/val, and no k/v is stored for
it, but an __IN_<key>/<val> is still kept so we still have access to the
original.
At MPI_*_set_info() time, opal_infosubscribe_change_info() is used. That
function calls the registered callbacks for each item in the provided info.
If the callback returns non-null, the info is updated with that k/v, or if
the callback returns null, that key is deleted from info. An __IN_<key>/<val>
is saved either way, and overwrites any previously saved value.
When MPI_*_get_info() is called, opal_info_dup_mpistandard() is used, which
allows relatively easy changes in interpretation of the standard, by looking
at both the <key>/<val> and __IN_<key>/<val> in info. Right now it does
1. includes system extras, eg k/v defaults not expliclty set by the user
2. omits ignored keys
3. shows input values, not callback modifications, eg not the internal values
Currently the callbacks are doing things like
return some_condition ? "true" : "false"
that is, returning static strings that are not to be freed. If the return
strings start becoming more dynamic in the future I don't see how unallocated
strings could support that, so I'd propose a change for the future that
the callback()s registered with info_subscribe_subscribe() do a strdup on
their return, and we change the callers of callback() to free the strings
it returns (there are only two callers).
Rough outline of the smaller changes spread over the less central files:
comm.c
initialize comm->super.s_info to NULL
copy into comm->super.s_info in comm creation calls that provide info
OBJ_RELEASE comm->super.s_info at free time
comm_init.c
initialize comm->super.s_info to NULL
file.c
copy into file->super.s_info if file creation provides info
OBJ_RELEASE file->super.s_info at free time
win.c
copy into win->super.s_info if win creation provides info
OBJ_RELEASE win->super.s_info at free time
comm_get_info.c
file_get_info.c
win_get_info.c
change_info() if there's no info attached (shouldn't happen if callbacks
are registered)
copy the info for the user
The other category of change is generally addressing compiler warnings where
ompi_info_t and opal_info_t were being used a little too interchangably. An
ompi_info_t* contains an opal_info_t*, at &(ompi_info->super)
Also this commit updates the copyrights.
Signed-off-by: Mark Allen <markalle@us.ibm.com>
ompi_communicator_t, ompi_win_t, ompi_file_t all have a super class of type opal_infosubscriber_t instead of a base/super type of opal_object_t (in previous code comm used c_base, but file used super). It may be a bit bold to say that being a subscriber of MPI_Info is the foundational piece that ties these three things together, but if you object, then I would prefer to turn infosubscriber into a more general name that encompasses other common features rather than create a different super class. The key here is that we want to be able to pass comm, win and file objects as if they were opal_infosubscriber_t, so that one routine can heandle all 3 types of objects being passed to it.
MPI_INFO_NULL is still an ompi_predefined_info_t type since an MPI_Info is part of ompi but the internal details of the underlying information concept is part of opal.
An ompi_info_t type still exists for exposure to the user, but it is simply a wrapper for the opal object.
Routines such as ompi_info_dup, etc have all been moved to opal_info_dup and related to the opal directory.
Fortran to C translation tables are only used for MPI_Info that is exposed to the application and are therefore part of the ompi_info_t and not the opal_info_t
The data structure changes are primarily in the following files:
communicator/communicator.h
ompi/info/info.h
ompi/win/win.h
ompi/file/file.h
The following new files were created:
opal/util/info.h
opal/util/info.c
opal/util/info_subscriber.h
opal/util/info_subscriber.c
This infosubscriber concept is that communicators, files and windows can have subscribers that subscribe to any changes in the info associated with the comm/file/window. When xxx_set_info is called, the new info is presented to each subscriber who can modify the info in any way they want. The new value is presented to the next subscriber and so on until all subscribers have had a chance to modify the value. Therefore, the order of subscribers can make a difference but we hope that there is generally only one subscriber that cares or modifies any given key/value pair. The final info is then stored and returned by a call to xxx_get_info.
The new model can be seen in the following files:
ompi/mpi/c/comm_get_info.c
ompi/mpi/c/comm_set_info.c
ompi/mpi/c/file_get_info.c
ompi/mpi/c/file_set_info.c
ompi/mpi/c/win_get_info.c
ompi/mpi/c/win_set_info.c
The current subscribers where changed as follows:
mca/io/ompio/io_ompio_file_open.c
mca/io/ompio/io_ompio_module.c
mca/osc/rmda/osc_rdma_component.c (This one actually subscribes to "no_locks")
mca/osc/sm/osc_sm_component.c (This one actually subscribes to "blocking_fence" and "alloc_shared_contig")
Signed-off-by: Mark Allen <markalle@us.ibm.com>
Conflicts:
AUTHORS
ompi/communicator/comm.c
ompi/debuggers/ompi_mpihandles_dll.c
ompi/file/file.c
ompi/file/file.h
ompi/info/info.c
ompi/mca/io/ompio/io_ompio.h
ompi/mca/io/ompio/io_ompio_file_open.c
ompi/mca/io/ompio/io_ompio_file_set_view.c
ompi/mca/osc/pt2pt/osc_pt2pt.h
ompi/mca/sharedfp/addproc/sharedfp_addproc.h
ompi/mca/sharedfp/addproc/sharedfp_addproc_file_open.c
ompi/mca/topo/treematch/topo_treematch_dist_graph_create.c
ompi/mpi/c/lookup_name.c
ompi/mpi/c/publish_name.c
ompi/mpi/c/unpublish_name.c
opal/mca/mpool/base/mpool_base_alloc.c
opal/util/Makefile.am
One should use the correct module object when calling
c_coll.coll_allgather. Otherwise there will be a segfault in the
case, for example, when hcoll is used. In that case
c_coll.coll_allgather = mca_coll_hcoll_allgather while
c_coll.coll_gather_module = tuned.
Signed-off-by: Valentin Petrov <valentinp@mellanox.com>
As we changed the ABI (forcing a major release), we can limit
the size of the predefined communicators by moving the collective
structure outside the communicator. This might have a minimal,
but unnoticeable, impact on performance. This approach has been
discussed during the January 2017 devel meeting.
Signed-off-by: George Bosilca <bosilca@icl.utk.edu>
Signed-off-by: Joshua Hursey <jhursey@us.ibm.com>
This commit adds support for using network AMOs for MPI_Accumulate,
MPI_Fetch_and_op, and MPI_Compare_and_swap. This support is only
enabled if the ompi_single_intrinsic info key is specified or the
acc_single_interinsic MCA variable is set. This configuration
indicates to this implementation that no long accumulates will be
performed since these do not currently mix with the AMO
implementation.
This commit also cleans up the code somwhat. This includes removing
unnecessary struct keywords where the type is also typedef'd.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
Need to increment the total size after checking the local offset not
before. This typo causes large allocations with MPI_Win_allocate() to
fail.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit fixes a bug that occurs when ranks are either not mapped
evenly or by something other than core.
Fixesopen-mpi/ompi#1599
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
If atomics are not globally visible (cpu and nic atomics do not mix)
then a btl endpoint must be used to access local ranks. To avoid
issues that are caused by having the same region registered with
multiple handles osc/rdma was updated to always use the handle for
rank 0. There was a bug in the update that caused osc/rdma to continue
using the local endpoint for accessing the state even though the
pointer/handle are not valid for that endpoint. This commit fixes the
bug.
Fixesopen-mpi/ompi#1241.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
Optimizing put aggregation in the presence of threads will require a
redesign of the code. For now just ensure that put aggregation is
turned off when MPI_THREAD_MULTIPLE is enabled.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
NOTE: Building with external pmix *requires* that you also build with external libevent and hwloc libraries. Detect this at configure and error out with large message if this requirement is violated.
Closes#1204 (replaces it)
Fixes#1064
There were two bugs in osc/rdma when using threads:
- Deadlock is ompi_osc_rdma_start_atomic. This occurs because
ompi_osc_rdma_frag_alloc is called with the module lock. To fix the
issue the module lock is now recursive. In the future I will add a
new lock to protect just the current rdma fragment.
- Do not drop the lock in ompi_osc_rdma_frag_alloc when calling
ompi_osc_rdma_frag_complete. Not only is it not needed but dropping
the lock at this point can cause a competing thread to mess up the
state.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit fixes a bug that can occur on Cray Gemini networks. If
multiple registrations are used for the local state then we looks the
atomicity guarantees. To avoid issues like this use only a single
registration handle for all local state on a node.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit fixes several bugs in the osc/rdma component:
- Complete aggregated requests immediately. Completion of RMA
requests indicates local completion anyway. This fixes a hang in
the c_reqops test.
- Correctly mark Rget_accumulate requests.
- Set the local base flag correctly on the local peer.
- Clear or set the no locks flag on the window if the value is
changed by MPI_Win_set_info.
- Actually update the target when using MPI_OP_REPLACE.
Fixesopen-mpi/ompi#1010
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit fixes an issue identified by @rolfv. The local peer was
not being correctly initialized when running with a single process on
a node.
This fixesopen-mpi/ompi#1010
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit adds support for performing one-sided operations over
supported hardware (currently Infiniband and Cray Gemini/Aries). This
component is still undergoing active development.
Current features:
- Use network atomic operations (fadd, cswap) for implementing
locking and PSCW synchronization.
- Aggregate small contiguous puts.
- Reduced memory footprint by storing window data (pointer, keys,
etc) at the lowest rank on each node. The data is fetched as each
process needs to communicate with a new peer. This is a trade-off
between the performance of the first operation on a peer and the
memory utilization of a window.
TODO:
- Add support for the accumulate_ops info key. If it is known that
the same op or same op/no op is used it may be possible to use
hardware atomics for fetch-and-op and compare-and-swap.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL
All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic.
This commit was SVN r32317.
The post and start window calls are supposed to be matching. The code
did not check to see that an incoming post matched with the start call.
This commit fixes the bug by placing the post on a pending list that
will be checked by the next call to start.
cmr=v1.8.2:reviewer=dgoodell
This commit was SVN r32017.
This commit fixes a bug that can cause request and communicator leaks
when cleaning up an OSC window. The should prevent a hang seen with
IMB-EXT.
cmr=v1.8.2:reviewer=jsquyres
This commit was SVN r31539.
This commit adds large datatype description support to the osc/rdma
component. Support is provided by an additional send/recv of the datatype
description if the description does not fit in an eager buffer. The
code is designed to require minimal new code and not for speed. We
consider this code path to be a slow path.
Refs trac:1905
cmr=v1.8:reviewer=jsquyres
This commit was SVN r31197.
The following Trac tickets were found above:
Ticket 1905 --> https://svn.open-mpi.org/trac/ompi/ticket/1905
It seems we can't release accumulate buffers in completion callbacks
because the btls don't release registration resources until after the
callback has fired. The fix is to keep track of the unused buffers and
free them later. This should resolve issues when running IMB-EXT and
IMB-RMA.
cmr=v1.7.5:reviewer=jsquyres
This commit was SVN r31029.
