The direct modex operation is slow, especially at scale for even modestly-connected applications. Likewise, blocking in MPI_Init while we wait for a full modex to complete takes too long. However, as George pointed out, there is a middle ground here. We could kickoff the modex operation in the background, and then trap any modex_recv's until the modex completes and the data is delivered. For most non-benchmark apps, this may prove to be the best of the available options as they are likely to perform other (non-communicating) setup operations after MPI_Init, and so there is a reasonable chance that the modex will actually be done before the first modex_recv gets called.
Once we get instant-on-enabled hardware, this won't be necessary. Clearly, zero time will always out-perform the time spent doing a modex. However, this provides a decent compromise in the interim.
This PR changes the default settings of a few relevant params to make "background modex" the default behavior:
* pmix_base_async_modex -> defaults to true
* pmix_base_collect_data -> continues to default to true (no change)
* async_mpi_init - defaults to true. Note that the prior code attempted to base the default setting of this value on the setting of pmix_base_async_modex. Unfortunately, the pmix value isn't set prior to setting async_mpi_init, and so that attempt failed to accomplish anything.
The logic in MPI_Init is:
* if async_modex AND collect_data are set, AND we have a non-blocking fence available, then we execute the background modex operation
* if async_modex is set, but collect_data is false, then we simply skip the modex entirely - no fence is performed
* if async_modex is not set, then we block until the fence completes (regardless of collecting data or not)
* if we do NOT have a non-blocking fence (e.g., we are not using PMIx), then we always perform the full blocking modex operation.
* if we do perform the background modex, and the user requested the barrier be performed at the end of MPI_Init, then we check to see if the modex has completed when we reach that point. If it has, then we execute the barrier. However, if the modex has NOT completed, then we block until the modex does complete and skip the extra barrier. So we never perform two barriers in that case.
HTH
Ralph
Signed-off-by: Ralph Castain <rhc@open-mpi.org>
The usNIC BTL does not use more than 1 iov, so be sure to set it to 1
so that we don't allocate cq/rq/sq entries based on a default (i.e.,
>1) number of iovs per entry.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
This PR renames the common library for OFI libfabric from
libfabric to ofi. There are a number of reasons this
is good to do:
1) its shorter and replaces 9 characters with three for
function names for what may eventually be a fairly extensive interface
2) OFI is the term used for MTL and RML components that use
the OFI libfabric interface
3) A planned OSC component will also use the OFI term.
4) Other HPC libraries that can use OFI libfabric tend to use
the term "ofi" internally and also in their configure options
relevant to OFI libfabric (i.e. MPICH/CH4, Intel MPI, Sandia SHMEM)
There seem to be comments in places in the Open MPI source
code that indicate that this common library will be going away.
Far from it as we will want to be able to share things like
AV objects between OMPI and possibly OSHMEM components that
use the OFI libfabric interface.
This PR also adds a synonym to the --with-libfabric(-libdir)
configury options: --with-ofi and with-ofi-libdir.
Signed-off-by: Howard Pritchard <howardp@lanl.gov>
This commit recategorizes several mpirun arguments,
and moves the information for mpirun --help arguments
to the bottom of the general help message. I also
added the OPAL_CMD_LINE_OTYPE field to two commands
that were missed initially because they were not
in the same area as the others.
Signed-off-by: Nathaniel Graham <ngraham@lanl.gov>
With this, libs (e.g., "-ldl") are not added to the wrapper LIBS
flags. This may work on some platforms, but on at least RHEL 7.3, it
does not (i.e., compiling MPI applications fails because it can't find
dlopen).
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Set the daemons' state to "running" and mark them as "alive" by default when constructing the nidmap
Get the DVM running again
Fix direct modex by eliminating race condition caused by releasing data while sending it
Up the size limit before compressing
Signed-off-by: Ralph Castain <rhc@open-mpi.org>
Commit fec519a793 broke the ability to
run autogen.pl in a distribution tarball. This commit restores that
ability by also distributing opal/mca/hwloc/autogen.options in the
tarball.
Skipping CI because CI does not test this functionality:
[skip ci]
bot:notest
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Several component-specific functions were named with a prefix of
"opal_timer_base", which was quite confusing. Rename them to have a
prefix "opal_timer_linux" to make it clear that they are here in this
component (and different than *actual* opal_timer_base symbols).
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
This commit makes datagram checks time based and reduces their
frequency when only the wildcard datagram is posted. This change
improves latency on knl systems.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit removes the local operation count check from the local SMSG
completion queue. This check was leading to hangs due to an undocumented
feature of the ugni library. The local SMSG CQ is used to send credit
return messages back to the sender. The ugni library never checks for
the completion itself but relying on the SMSG user to periodically
check the CQ.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit updates the ugni btl to make use of multiple device
contexts to improve the multi-threaded RMA performance. This commit
contains the following:
- Cleanup the endpoint structure by removing unnecessary field. The
structure now also contains all the fields originally handled by the
common/ugni endpoint.
- Clean up the fragment allocation code to remove the need to
initialize the my_list member of the fragment structure. This
member is not initialized by the free list initializer function.
- Remove the (now unused) common/ugni component. btl/ugni no longer
need the component. common/ugni was originally split out of
btl/ugni to support bcol/ugni. As that component exists there is no
reason to keep this component.
- Create wrappers for the ugni functionality required by
btl/ugni. This was done to ease supporting multiple device
contexts. The wrappers are thread safe and currently use a spin
lock instead of a mutex. This produces better performance when
using multiple threads spread over multiple cores. In the future
this lock may be replaced by another serialization mechanism. The
wrappers are located in a new file: btl_ugni_device.h.
- Remove unnecessary device locking from serial parts of the ugni
btl. This includes the first add-procs and module finalize.
- Clean up fragment wait list code by moving enqueue into common
function.
- Expose the communication domain flags as an MCA variable. The
defaults have been updated to reflect the recommended setting for
knl and haswell.
- Avoid allocating fragments for communication with already
overloaded peers.
- Allocate RDMA endpoints dyncamically. This is needed to support
spreading RMA operations accross multiple contexts.
- Add support for spreading RMA communication over multiple ugni
device contexts. This should greatly improve the threading
performance when communicating with multiple peers. By default the
number of virtual devices depends on 1) whether
opal_using_threads() is set, 2) how many local processes are in the
job, and 3) how many bits are available in the pid. The last is
used to ensure that each CDM is created with a unique id.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit exposes ugni statistics for use with MPI_T. There is
no overhead to providing these counters.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
