attempting to break the get into multiple rdma fragments
A little background. Historically ob1 always registered the entire memory
region when the RGET protocol was in use. This changed when Mellanox
added support to fragment RGET using the btl_prepare_dst function. Now
that the BTL layer has changed to split out the limits of get/put there
is explicit fragmentation code in ob1. Before this commit the registration
was still done per RGET fragment.
This commit will attempt to register the entire region before creating
RGET fragments. If the registration is successfull then all RGET
fragments will use this registration otherwise they will each attempt
to register their own segment of the receive buffer. If that fails
enough times each fragment will give up and fall back on send/recv.
opal_mutex_t must be OBJ_DESTRUCTed in order to avoid
a memory leak (pthread_mutex_init allocates memory under
Cygwin, so pthread_mutex_destroy is mandatory)
Thanks to Marco Atzeri for reporting this issue
These two macros set the prefix for the OPAL and ORTE libraries,
respectively. Specifically, the OPAL library will be named
libPREFIXopen-pal.la and the ORTE library will be named
libPREFIXopen-rte.la.
These macros must be called, even if the prefix argument is empty.
The intent is that Open MPI will call these macros with an empty
prefix, but other projects (such as ORCM) will call these macros with
a non-empty prefix. For example, ORCM libraries can be named
liborcm-open-pal.la and liborcm-open-rte.la.
This scheme is necessary to allow running Open MPI applications under
systems that use their own versions of ORTE and OPAL. For example,
when running MPI applications under ORTE, if the ORTE and OPAL
libraries between OMPI and ORCM are not identical (which, because they
are released at different times, are likely to be different), we need
to ensure that the OMPI applications link against their ORTE and OPAL
libraries, but the ORCM executables link against their ORTE and OPAL
libraries.
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 wrong descriptor field was used when calculating the size received when
using the RDMA rendevous protcol.
This commit was SVN r32232.
The following SVN revision numbers were found above:
r32196 --> open-mpi/ompi@a14e0f10d4
Handle OMPI_REQUEST_NOOP in MPI_Startall rather than PML
cmr=v1.8.2:reviewer=bosilca:ticket=4764
This commit was SVN r32213.
The following Trac tickets were found above:
Ticket 4764 --> https://svn.open-mpi.org/trac/ompi/ticket/4764
mca_btl_base_segment_t and replace them with des_local and des_remote
This change also updates the BTL version to 3.0.0. This commit does
not represent the final version of BTL 3.0.0. More changes are coming.
In making this change I updated all of the BTLs as well as BTL user's
to use the new structure members. Please evaluate your component to
ensure the changes are correct.
RFC text:
This is the first of several BTL interface changes I am proposing for
the 1.9/2.0 release series.
What: Change naming of btl descriptor members. I propose we change
des_src and des_dst (and their associated counts) to be des_local and
des_remote. For receive callbacks the des_local member will be used to
communicate the segment information to the callback. The proposed change
will include updating all of the doxygen in btl.h as well as updating
all BTLs and BTL users to use the new naming scheme.
Why: My btl usage makes use of both put and get operations on the same
descriptor. With the current naming scheme I need to ensure that there
is consistency beteen the segments described in des_src and des_dst
depending on whether a put or get operation is executed. Additionally,
the current naming prevents BTLs that do not require prepare/RMA matched
operations (do not set MCA_BTL_FLAGS_RDMA_MATCHED) from executing
multiple simultaneous put AND get operations. At the moment the
descriptor can only be used with one or the other. The naming change
makes it easier for BTL users to setup/modify descriptors for RMA
operations as the local segment and remote segment are always in the
same member field. The only issue I forsee with this change is that it
will require a little more work to move BTL fixes to the 1.8 release
series.
This commit was SVN r32196.
This corner case is now handled in the pml so the same code
is invoked for both MPI_Start and MPI_Startall.
This also correctly report an error if MPI_Startall is invoked twice
on a MPI_PROC_NULL persistent request.
This commit was SVN r32139.
Based on a suggestion from George on #31806, use the pre-computed
sizes rather than duplicating the computation math (which may change
someday in the future).
cmr=v1.8.2:ticket=trac:4647
This commit was SVN r31841.
The following Trac tickets were found above:
Ticket 4647 --> https://svn.open-mpi.org/trac/ompi/ticket/4647
a memory leak. Similar changes shuld be applied to all the
other PML that are copies of OB1. This patch is related to
#4653.
This commit was SVN r31838.
MPI_Isend.
I filed an RFC for this optimization some time back. It is a
relatively simple optimization. If the data associated with an
MPI_Isend can be put on the wire without allocating an MPI_Request
then do so. In this case we can legally return omp_request_empty
which will correctly indicate that the request is complete and that is
was not cancelled (these are the only requirements on send requests).
cmr=v1.8.3:reviewer=bosilca
This commit was SVN r31828.
r30343 introduced the optimization of putting the OB1 sendreq and
recvreq on the stack for blocking sends and receives. However, the
requests did not contain enough storage for the data that is normally
immediately ''after'' the request (e.g., BTL data).
This commit changes these requests to be pointers and to use alloca()
to get enough total space for the OB1 request and all the associated
data.
The change is smaller than it looks; most of it is just changing from
"foo.bar" to "foo->bar" notation (etc.).
Submitted by Jeff, reviewed by Nathan. But we want George to look at
this (and get a little soak time on the trunk) before moving to v1.8.
cmr=v1.8.2:reviewer=bosilca
This commit was SVN r31806.
The following SVN revision numbers were found above:
r30343 --> open-mpi/ompi@2b57f4227e
An OBJ_DESTRUCT was missing for mca_pml_ob1.send_ranges causing a
memory leak. Identified by valgrind.
cmr=v1.8.2:reviewer=jsquyres
This commit was SVN r31768.
top_ompi_srcdir -> OMPI_TOP_SRCDIR
top_ompi_builddir -> OMPI_TOP_BUILDDIR
We also split the srcdir/builddir flags according to their local tree (e.g., OPAL_TOP_SRCDIR), and tied them all together in configure.ac. Renamed ompi_ignore and ompi_unignore to be opal_<foo> as these are agnostic markers.
Only thing left is ompilibdir being treated similar to what we dif for srcdir/builddir. Coming soon.
This commit was SVN r31678.
We will track #4568 from the 1.8 CMR.
Closes trac:4568
cmr=v1.8.2:reviewer=jsquyres
This commit was SVN r31535.
The following Trac tickets were found above:
Ticket 4568 --> https://svn.open-mpi.org/trac/ompi/ticket/4568
The problem was caused by the static request optimization. The buffered send case
is much like the isend case in that the request structure may be needed after
MPI_Bsend completes. Fix this case by calling isend and freeing the resulting
request.
cmr=v1.7.5:ticket=trac:4149
This commit was SVN r30601.
The following Trac tickets were found above:
Ticket 4149 --> https://svn.open-mpi.org/trac/ompi/ticket/4149
This change does not appear to increase the small message latency of ping-pong
benchmarks and fixes an issue found by our ibm datatype tests.
Fixes trac:4232
cmr=v1.7.5:ticket=trac:4149
This commit was SVN r30598.
The following Trac tickets were found above:
Ticket 4149 --> https://svn.open-mpi.org/trac/ompi/ticket/4149
Ticket 4232 --> https://svn.open-mpi.org/trac/ompi/ticket/4232
btl sendi functions currently can not handle the descriptor being NULL. The
send inline optimization was assuming (incorrectly) that NULL was ok.
cmr=v1.7.5:ticket=trac:4149
This commit was SVN r30364.
The following Trac tickets were found above:
Ticket 4149 --> https://svn.open-mpi.org/trac/ompi/ticket/4149
Per RFC. There are two optimizations in this commit:
- Allocate requests for blocking sends and receives on the stack. This
bypasses the request free list and saves two atomics on the critical path.
This change improves the small message ping-pong by 50-200ns on both AMD
and Intel CPUs.
- For small messages try to use the btl sendi function before intializing a
send request. If the sendi fails or the btl does not have a sendi function
silently fallback on the standard send path.
cmr=v1.7.5:reviewer=brbarret
This commit was SVN r30343.
pkg{data,lib,includedir}, use our own ompi{data,lib,includedir}, which is
always set to {datadir,libdir,includedir}/openmpi. This will keep us from
having help files in prefix/share/open-rte when building without Open MPI,
but in prefix/share/openmpi when building with Open MPI.
This commit was SVN r30140.
http://www.open-mpi.org/community/lists/devel/2013/10/13072.php
Add support for pinning GPU Direct RDMA in openib BTL for better small message latency of GPU buffers.
Note that none of this is compiled in unless CUDA-aware support is requested.
This commit was SVN r29680.
configure-time dynamic allocation of flags. The net result for platforms
which only support BTL-based communication is a reduction of 8*nprocs bytes
per process. Platforms which support both MTLs and BTLs will not see
a space reduction, but will now be able to safely run both the MTL and BTL
side-by-side, which will prove useful.
This commit was SVN r29100.
*** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE ***
Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro.
***************************************************************************************
I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week.
The code is in https://bitbucket.org/rhc/ompi-oob2
WHAT: Rewrite of ORTE OOB
WHY: Support asynchronous progress and a host of other features
WHEN: Wed, August 21
SYNOPSIS:
The current OOB has served us well, but a number of limitations have been identified over the years. Specifically:
* it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code)
* we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface.
* the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients
* there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort
* only one transport (i.e., component) can be "active"
The revised OOB resolves these problems:
* async progress is used for all application processes, with the progress thread blocking in the event library
* each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on")
* multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC.
* a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions.
* opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object
* NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions
* obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel
* the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport
* routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active
* all blocking send/recv APIs have been removed. Everything operates asynchronously.
KNOWN LIMITATIONS:
* although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline
* the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker
* routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways
* obviously, not every error path has been tested nor necessarily covered
* determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost.
* reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways
* the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC
This commit was SVN r29058.
