1
1
Граф коммитов

1376 Коммитов

Автор SHA1 Сообщение Дата
Brian Barrett
6f8b366acb Rename liborte to libopen-rte and libopal to libopen-pal per telecon today
and bug #632.

Refs trac:632

This commit was SVN r12762.

The following Trac tickets were found above:
  Ticket 632 --> https://svn.open-mpi.org/trac/ompi/ticket/632
2006-12-05 18:27:24 +00:00
Brian Barrett
441432950f Merge in changes from the bwb-heterogeneous temp branch (r12491 -
r12714) for supporting compilers / architectures with different
padding rules.

This commit was SVN r12749.

The following SVN revisions from the original message are invalid or
inconsistent and therefore were not cross-referenced:
  r12491
  r12714
2006-12-04 20:11:42 +00:00
Rainer Keller
6f8f28f40f - Get rid of inline definition, otherwise static-compilation fails.
This commit was SVN r12735.
2006-12-03 14:52:17 +00:00
Gleb Natapov
30ca7457b4 Some BTLs (e.g TCP) can report put/get completion before data actually
hits the buffer on the other side. For this kind of BTLs we need to send
FIN through the same BTL, PUT was performed with so network will handle
ordering for us. If we will use another BTL, receiver can get FIN before
data will hit the buffer and complete request prematurely. We mark such
problematic BTLs with MCA_BTL_FLAGS_FAKE_RDMA flag (this kind of RDMA
is really fake, because the real one guaranties that sender will see the
completion only after receiver's NIC confirmed that all the data was
received).

This commit was SVN r12732.
2006-12-03 10:12:09 +00:00
Gleb Natapov
39c930b160 The bug fixing part of r12720 introduce much more serious bug that it fixes.
It calls mca_pml_ob1_send_fin_btl() which may fail and doesn't check return
code. This breaks all RDMA transports event when only one BTL is used. Revert
it for now, I am working on a real fix for the problem (I hope).

This commit was SVN r12731.

The following SVN revision numbers were found above:
  r12720 --> open-mpi/ompi@3e3689320b
2006-12-03 08:55:59 +00:00
Gleb Natapov
65d7ad4581 The "bug fix" from the r12721 reverts part of the r12433 that fixed
regresion from v1.1 was reviewed and put to v1.2 branch. So revert this part
of r12721 back.

This commit was SVN r12730.

The following SVN revision numbers were found above:
  r12433 --> open-mpi/ompi@82f7c0dd69
  r12721 --> open-mpi/ompi@3edd850d2e
2006-12-03 08:29:55 +00:00
George Bosilca
3edd850d2e Some indentation and code arrangement. However, there is a bug fix. Force the PUT
protocol to always obey to the btl_max_rdma_size.

This commit was SVN r12721.
2006-12-01 22:26:14 +00:00
George Bosilca
3e3689320b Some indentations and one BIG fix. Avoid race conditions on the PUT RDMA
protocol when multiple NICS are available between 2 peers. The fix force
the FIN message to take exactly the same path as the fragment it describe
(i.e. same path means same BTL). Otherwise, the FIN can be received by
the peer before the RDMA complete and the request will get freed
too early.

This commit was SVN r12720.
2006-12-01 21:52:07 +00:00
George Bosilca
658879232b Several small improvements:
- consistent error message when something fails (via BTL_ERROR macro)
- decrease the number of jumps.
- cleanup some parts of the code.

This commit was SVN r12719.
2006-12-01 21:48:06 +00:00
Brian Barrett
2a09fa2d9d * silence compiler warning
This commit was SVN r12717.
2006-12-01 20:01:53 +00:00
Brian Barrett
bfbc281e93 Fix slow startup issue with the MX MTL. The problem is caused by mx_connect() being a one-sided operation from the API level, but not being an interrupting call when the target is not entering the MX library. So if most of the processes exit mtl_mx_add_procs() and enter the stage gate 2 barrier, the other processes can only progress their mx_connect() calls when the targets enter the mx library. Because the event library is in EV_ONELOOP mode, this only happens once a second. The mx progress thread (hidden in the MX library) also only wakes up once a second, so mx_connect calls can take a second to complete.
The temporary solution is to switch into EV_NONBLOCK mode earlier (right after the mx_connect loop) so that there isn't a giant slowdown when processes enter the stage gate 2 barrier before other proesses.  They will now not block in the event library for any period of time, which appears to have a 50% speedup when running at > 64 procs.

Refs trac:645

This commit was SVN r12713.

The following Trac tickets were found above:
  Ticket 645 --> https://svn.open-mpi.org/trac/ompi/ticket/645
2006-12-01 02:49:01 +00:00
Pavel Shamis
f08bc818c4 Cleaning mca_btl_openib_progress_thread from
unused variables.

This commit was SVN r12709.
2006-11-30 18:28:45 +00:00
Brian Barrett
bc3250fe6f Fix issue where messages could start arriving before the window was
fully initialized, especially during lock/unlock, which doesn't use MPI
collectives for synchronization (unlike Fence)

This commit was SVN r12676.
2006-11-27 22:42:21 +00:00
Brian Barrett
b5b6f4c4bf Remove check for epoch on the receiver side, as this is a race condition between message arrival, and because we'll be still blocking in a synchronization call, is of no consequence to the user. We do an epoch check on the sending side,
so this isn't an issue there either.  Refs trac:488

This commit was SVN r12675.

The following Trac tickets were found above:
  Ticket 488 --> https://svn.open-mpi.org/trac/ompi/ticket/488
2006-11-27 22:15:34 +00:00
Brian Barrett
beb1e9d4dd * finish move from hard coded tag to #define'd constant tag
This commit was SVN r12674.
2006-11-27 21:55:41 +00:00
Brian Barrett
0c25f7be09 More One-sided fixes:
* Fix a counter roll-over issue that could result from a large (but
    not excessive) number of outstanding put/get/accumulate calls
    during a single synchronization issues (Refs trac:506)
  * Fix epoch issue with rdma component that would effect PWSC
    synchronization (Refs trac:507)

This commit was SVN r12673.

The following Trac tickets were found above:
  Ticket 506 --> https://svn.open-mpi.org/trac/ompi/ticket/506
  Ticket 507 --> https://svn.open-mpi.org/trac/ompi/ticket/507
2006-11-27 21:41:29 +00:00
George Bosilca
59cfee0cd2 Use the MX infinite timeout by default. The user can modify it using an MCA
parameter.

This commit was SVN r12670.
2006-11-27 20:18:58 +00:00
Brian Barrett
993d2a7753 Fix for issue IU is seeing on BigRed with connections timing out during
MPI_INIT.  Use an infinite timeout, which is exactly what MPICH-MX does.

This commit was SVN r12669.
2006-11-27 20:10:27 +00:00
Brian Barrett
63e5668e29 Number of one-sided fixes:
* use one-sided datatype check instead of send/receive and check both
    the origin and target datatypes
  * allow error handler to be set on MPI_WIN_NULL, per standard
  * Allow recursive calls into the pt2pt osc component's progress
    function
  * Fix an uninitialized variable problem in the unlock header

This commit was SVN r12667.
2006-11-27 03:22:44 +00:00
Brian Barrett
0895f5e08d Rename OMPI_PROCESS_NAME_{HTON, NTOH} macros to ORTE_PROCESS_NAME_{HTON, NTOH}
because they are in ORTE, not OMPI.  Also, remove the ORTE_PROCESS_NAME macros
in iof base as they are duplicates of the ones that were in ns_types, which 
meant that bad things happened if you changed what an orte_process_name_t
looked like.

This commit was SVN r12646.
2006-11-22 03:03:21 +00:00
Brian Barrett
33320b7165 Rework the opal_progress interface to better support dynamic processes and at
the same time, remove some of the MPI-related options from OPAL:

  - provide mechanism to change at runtime whether sched_yield() should 
    be called when the progress engine is idle
  - provide mechanism for changing the rate at which the event engine
    is called when there are "no" users of the event engine (ie, when
    using MPI but not TCP)
  - fix some function names in the progress engine to better match
    their intended use (and remove MPI naming scheme)
  - remove progress_mpi_enable / progress_mpi_disable because 
    we can now use the functions to set the sched_yield and
    tick rate interfaces
  - rename opal_progress_events() to opal_progress_set_event_flag()
    because the first really isn't descriptive of what the function
    does and I always got confused by it

This commit was SVN r12645.
2006-11-22 02:06:52 +00:00
Ralph Castain
6d6cebb4a7 Bring over the update to terminate orteds that are generated by a dynamic spawn such as comm_spawn. This introduces the concept of a job "family" - i.e., jobs that have a parent/child relationship. Comm_spawn'ed jobs have a parent (the one that spawned them). We track that relationship throughout the lineage - i.e., if a comm_spawned job in turn calls comm_spawn, then it has a parent (the one that spawned it) and a "root" job (the original job that started things).
Accordingly, there are new APIs to the name service to support the ability to get a job's parent, root, immediate children, and all its descendants. In addition, the terminate_job, terminate_orted, and signal_job APIs for the PLS have been modified to accept attributes that define the extent of their actions. For example, doing a "terminate_job" with an attribute of ORTE_NS_INCLUDE_DESCENDANTS will terminate the given jobid AND all jobs that descended from it.

I have tested this capability on a MacBook under rsh, Odin under SLURM, and LANL's Flash (bproc). It worked successfully on non-MPI jobs (both simple and including a spawn), and MPI jobs (again, both simple and with a spawn).

This commit was SVN r12597.
2006-11-14 19:34:59 +00:00
George Bosilca
139f9cf3d0 Make sure we disable the MX shared memory when we use the MX BTL.
This commit was SVN r12587.
2006-11-13 22:17:06 +00:00
Andrew Friedley
a4bdcb4faa Fix a segfault that turned up in more MPI_THREAD_MULTIPLE testing.
Same sort of problem and fix as described in r12323 - mca_pml_ob1_recv_frag_progress() was segfaulting due to a NULL req_proc pointer.  The path leading to this was through the mca_pml_ob1_check_cantmatch_for_match() function, where we can match a frag using the same macros as mca_pml_ob1_frag_match() and never initialize the req_proc pointer.

This commit was SVN r12582.

The following SVN revision numbers were found above:
  r12323 --> open-mpi/ompi@c752502dee
2006-11-13 20:12:51 +00:00
Gleb Natapov
9933a6f469 Previous fix doesn't fix the case when opcode is changed in put/get functions.
The fix is to set opcode to SEND at the entrance to the send function before
checking credits and putting fragment to the pending list. We do the same thing
in put/get functions i.e setting opcode at the entrance to the function.

This commit was SVN r12559.
2006-11-11 07:51:06 +00:00
George Bosilca
ec410644ce Implement the send receive as 2 non blocking operations. That will help us
avoiding too many calls to opal_progress.

This commit was SVN r12553.
2006-11-10 23:06:19 +00:00
George Bosilca
c2c6a1b37e Correctly compute the number of elements in a segment.
For broadcast send the correct size for all intermediary nodes.

This commit was SVN r12552.
2006-11-10 23:04:50 +00:00
George Bosilca
7102147b9f Correctly detect when the specified algorithm is out of range. In
this case we reset it to zero.

This commit was SVN r12551.
2006-11-10 21:47:07 +00:00
George Bosilca
bfbd0e61f6 Minimize the number of lines of code :)
This commit was SVN r12550.
2006-11-10 20:56:08 +00:00
George Bosilca
a38cd366d7 Construct the convertor. It's not really required, but it's not in the
critical path anyway. At least in debug mode we get nice informations about
where the convertor was created.

This commit was SVN r12549.
2006-11-10 20:55:06 +00:00
George Bosilca
858ab24e8e The req_mtl field has to be the last in the struct or bad things happen.
This commit was SVN r12548.
2006-11-10 20:53:41 +00:00
George Bosilca
af68171253 Use the macro to compute the number of elements in a segment in both
bcast and reduce and update the default values for the variables
as required by the comment in the coll_tuned.h file.

This commit was SVN r12546.
2006-11-10 20:04:08 +00:00
George Bosilca
476b922074 Updates & upgrades:
- consistent arguments checking (not allowing to select an algorithm which
     is not available)
 - consistent way of computing the segcount (number of datatypes by segment).
 - small cleanups.
 - more informative debugging messages.

This commit was SVN r12545.
2006-11-10 19:54:09 +00:00
Gleb Natapov
7e03b83d23 Reset opcode field to SEND. It is checked later in pending progress function.
This commit was SVN r12531.
2006-11-10 06:17:00 +00:00
George Bosilca
77ef979457 New architecture for broadcast. A generic broadcast working on a tree
description. Most of the bcast algorithms can be completed using this
generic function once we create the tree structure. Add all kind of
trees.

There are 2 versions of the generic bcast function. One using overlapping
between receives (for intermediary nodes) and then blocking sends to all
childs and another where all sends are non blocking. I still have to
figure out which one give the smallest overhead.

This commit was SVN r12530.
2006-11-10 05:53:50 +00:00
George Bosilca
17405cd9c6 A temporary fix, until we figure out a better approach. The problem
is that if one add "pml=" to the configuration file, really bad things
happen. All PMLs will get initialize, and each of them will initialize
all BTLs. This patch force the mca_pml_base_pml to get initialized in
all cases before we go out of the mca_pml_base_open function.

This commit was SVN r12527.
2006-11-10 04:53:00 +00:00
George Bosilca
1d80f685b5 Remove one compiler warning.
This commit was SVN r12520.
2006-11-09 20:08:43 +00:00
George Bosilca
73eec4bfef Show the MCA parameter coll_base_verbose only if Open MPI is compiled in
debug mode. Otherwise there is no debug anyway ...

This commit was SVN r12516.
2006-11-09 19:02:32 +00:00
George Bosilca
a82ce427e4 Update the number of reduce algorithms available.
This commit was SVN r12503.
2006-11-08 22:20:34 +00:00
George Bosilca
eab1776e9a Explicit casts for our friendly Windows environment...
This commit was SVN r12496.
2006-11-08 17:02:46 +00:00
George Bosilca
0914892044 Small cleanups, some explicit casts.
This commit was SVN r12494.
2006-11-08 16:54:03 +00:00
George Bosilca
74d3946342 Remove the call to set_args. This is only required for the MPI level,
because there we have to be able to return to the user the description
of the data.

This commit was SVN r12493.
2006-11-08 16:52:48 +00:00
George Bosilca
915d748d72 Initialize the convertor on _START not on _INIT. This allow us to
set it up before the match when we know the peer, saving some
time on the critical path. If the receive is ANY_SOURCE then
we initialize the convertor on _MATCHED. Anyway, we will set it
up only once per receive.

This commit was SVN r12484.
2006-11-08 05:42:29 +00:00
George Bosilca
eb45a5e402 Move things around a little bit. Mainly fields from the send and receive
request in the base request. Rearrange the fields to keep the data
together. Remove some useless tests.

This commit was SVN r12482.
2006-11-08 04:58:23 +00:00
George Bosilca
63462331c9 Reduce the number of branches. Keep the fast path as short as possible.
Remove some useless error checking. Add OPAL_UNLIKELY directives.

This commit was SVN r12477.
2006-11-07 23:59:32 +00:00
George Bosilca
f3de2e1a82 Keep the fast path as short as possible.
This commit was SVN r12476.
2006-11-07 23:56:32 +00:00
Jeff Squyres
427c20af0d Use a new algorithm for allgatherv. The old algorithm essentially did
N gatherv's:

  for (i = 0 ... size)
    MPI_Gatherv(..., root = i, ...)

The new algorithm simply does (effectively):

  MPI_Gatherv(..., root = 0, ...)
  MPI_Bcast(..., root = 0, ...)

This commit was SVN r12469.
2006-11-07 18:07:55 +00:00
Galen Shipman
55db17b37c don't try to use a dead btl..
This commit was SVN r12456.
2006-11-06 23:25:24 +00:00
George Bosilca
108ea4dbe9 When the MX MTL complete a request, force a return from the progress function.
Decrease the latency by about 0.3 microseconds.

This commit was SVN r12454.
2006-11-06 23:13:07 +00:00
George Bosilca
3d0df2cf29 Allow the MX BTL to finish the small sends quicker. Once the mx_isend is posted if
the message size is less than 4K do a check for the message completion and if any
call the callback.

This commit was SVN r12453.
2006-11-06 23:12:01 +00:00