1
1
Граф коммитов

722 Коммитов

Автор SHA1 Сообщение Дата
George Bosilca
c1b713c56e Make a compiler happy about casting.
This commit was SVN r8005.
2005-11-07 04:59:46 +00:00
George Bosilca
7b7aaf897c Do not add epsilon to the data extent if there is a user set UB for the data.
This commit was SVN r8004.
2005-11-07 04:04:20 +00:00
Graham Fagg
dcd3450e06 simplified the building of different rule sets
(also corrected some prototypes missing 'struct')

This commit was SVN r8003.
2005-11-06 22:05:50 +00:00
Jeff Squyres
42ec26e640 Update the copyright notices for IU and UTK.
This commit was SVN r7999.
2005-11-05 19:57:48 +00:00
Jeff Squyres
2e10d0c099 Forgot to add the intrinsic op MPI_REPLACE (Brian, the One Sided Bug
Finder, gets credit)

This commit was SVN r7993.
2005-11-04 19:00:35 +00:00
Tim Woodall
e45f4744ee do not return these descriptors to cache
This commit was SVN r7986.
2005-11-03 23:20:38 +00:00
Tim Woodall
26003bc952 fix from release branch - don't use get protocol if more
than one btl is available

This commit was SVN r7984.
2005-11-03 20:52:56 +00:00
Jeff Squyres
9e723b8519 Remove some compiler warnings
This commit was SVN r7973.
2005-11-03 15:26:43 +00:00
Jeff Squyres
4fc135fd2b Looks like I forgot to put DDT support for the optional C datatypes
MPI_UNSIGNED_LONG_LONG, MPI_LONG_LONG, and MPI_LONG_LONG_INT --
although I already had implementations of all the relevant functions
for these types.  Doh!

This commit was SVN r7944.
2005-11-01 03:28:59 +00:00
Graham Fagg
9547a635a9 snapshot while switching systems
but, dynamic rules from a user defined config file is almost there now

This commit was SVN r7943.
2005-11-01 00:19:05 +00:00
Graham Fagg
fe03e068f2 allow forced algorithms (where the user or *test* suite knows better) to
go through the dynamic decision rule interface.
(forced algorithms are set with MCA params)
fixed some silly verbose output with wrong func name in it etc
updates to fixed dec rules.

This commit was SVN r7940.
2005-10-31 20:45:50 +00:00
Tim Woodall
31eb35c3f1 correct rnr parameter - need to review this code and pass correct data type
This commit was SVN r7936.
2005-10-31 17:18:39 +00:00
Edgar Gabriel
2ec5fa5d24 - The component will remove itself from the list of potential collective
modules, if its priority is zero (the default value). Reason for that is
  + if there is no other module with a priority > 0, the hierarchical
    collective module has a problem anyway, since it has to rely on the coll
    modules of the subcommunicators. On the other hand, if its priority is
    zero, it won't be chosen anyway, and we can simply save the
    allreduce/allgather and comm_split operations which might occur during
    hierarchy detection.
  + to improve the startup times until we have the modex thing which we
    discussed with Jeff and Tim in Knoxville in place

- adding an mca parameter indicating a symmetric configuration. This can 
  speed up startup times, since each process can conclude from its data onto
  the data of the other processes -> no need for the allreduce operations. Per
  default  this parameter is set to "no".

This commit was SVN r7932.
2005-10-30 16:01:13 +00:00
George Bosilca
b0def3f6bf MX has 2 limitations regarding the iovecs. First they do not support iovec witha total size
larger than 32K for inter-nodes transfert ... and then they do not support iovecs larger than
16K for inter-node transfert. Therefore we have to set the size of our first fragment to
16K to match both cases.

This commit was SVN r7926.
2005-10-28 20:37:43 +00:00
Jeff Squyres
7bdfe6557b - Update the checks in REDUCE, ALLREDUCE, SCAN, EXSCAN, and
REDUCE_SCATTER to more thoroughly check the datatype/op combination
  to see if it's valid or not.  If it's not, print a meaningful error
  message rather than "Invalid MPI_Op" indicating what specifically
  was wrong (therefore hopefully helping users track down where in the
  code the problem is, and/or telling us that there's a reduction
  operation combo that we don't support that we should)
- The check for whether a datatype is intrinsic needed to be updated
  -- it's not sufficient to check that dtype->id < DT_MAX_PREDEFINED;
  you really need to check the PREDEFINED flag on the datatype.
  Thanks to George for this fix (only intrinsics have a meaningful
  value in dtype->id).

This commit was SVN r7923.
2005-10-28 16:47:32 +00:00
George Bosilca
ab97bde177 Rainer pointer out that the convertor already have the CONTIGUOUS flag is the
data is contiguous (set in ompi_convertor_prepare).
For unpack reinforce the limits of the pack for contiguous types.

This commit was SVN r7914.
2005-10-28 05:27:40 +00:00
George Bosilca
5355765d81 Cleanup has to reset the stack position.
This commit was SVN r7913.
2005-10-28 05:25:08 +00:00
George Bosilca
d916e0c5b4 The (I hope) final solution for the convertor problem. As all the PML inherit
the base send and receive request from the pml_base, we can solve our problem
if we construct the convertor attached to any request in the pml_base_construct
function. At the end of the life time for each request (here life time is 
related to one utilisation, without taking in account the cache) we release
all information attached to the convertors in the _FINI macro by calling the
ompi_convertor_cleanup.

This commit was SVN r7910.
2005-10-28 03:26:36 +00:00
Brian Barrett
bf67c9387b * initialize send request convertor with the correct type (convertor instead
of request).  This fixes at least the bug with NetPIPE in 64bit land that
  Troy was seeing. 

This commit was SVN r7904.
2005-10-27 23:08:27 +00:00
Galen Shipman
4a15761732 add support for srq limit reached async event, even though it doesn't appear
to  be supported by mellanox vapi.. perhaps this will be supported in the near 
future, for now it doesn't hurt to have it in the trunk


Also cleanup the receive descriptor posting macro's.. 

This commit was SVN r7903.
2005-10-27 22:47:19 +00:00
Tim Woodall
3bd5b81dfa Submitted: Gleb Natapov
This commit was SVN r7899.
2005-10-27 17:48:40 +00:00
Tim Woodall
4fc5b2105a this is currently an int - we shouldn't restrict it unless required
This commit was SVN r7895.
2005-10-27 17:06:58 +00:00
Tim Woodall
13409ec53b correction for hang, check for additional fragments before callback,
which may queue a new fragment

This commit was SVN r7889.
2005-10-27 01:39:39 +00:00
Graham Fagg
5bb0d4a053 enable allreduce to be selected
This commit was SVN r7888.
2005-10-26 23:55:37 +00:00
Graham Fagg
2587d7ade9 added some more linear functions.
minor corrections on naming and debug info

This commit was SVN r7887.
2005-10-26 23:51:56 +00:00
Graham Fagg
c3e1dc410d Started to add basic linear functions
Also started to add the allreduce algorithms as I test them
(i.e. if it goes in its after testing from now on)

This commit was SVN r7886.
2005-10-26 23:11:32 +00:00
Jeff Squyres
d47ce065e9 Minor Makefile.am fix for static builds.
This commit was SVN r7882.
2005-10-26 15:57:58 +00:00
Edgar Gabriel
ba3bf6592f fixing some warnings. No idea yet why the static builds fail...
This commit was SVN r7879.
2005-10-26 12:56:56 +00:00
Jeff Squyres
23ab9e0277 A better solution to the previous commit -- RETAIN/RELEASE the MPI_Op
at the top-level MPI API function.  This allows two kinds of
scenarios:

1. MPI_Ireduce(..., op, ...);
   MPI_Op_free(op);
   MPI_Wait(...);

For the non-blocking collectives that we're someday planning -- to
make them analogous to non-blocking point-to-point stuff.

2. Thread 1:
   MPI_Reduce(..., op, ...);
   Thread 2:
   MPI_Op_free(op);

Granted, for #2 to occur would tread a fine line between a correct and
erroneous MPI program, but it is possible (as long as the Op_free was
*after* MPI_reduce() had started to execute).  It's more realistic
with case #1, where the Op_free() could be executed in the same thread
or a different thread.

This commit was SVN r7870.
2005-10-25 19:20:42 +00:00
Edgar Gabriel
d009d8de57 opening the hierarchical collective component to the public. I am at this
stage fairly confident that 
- it works in most scenarious (with symmetric hierarchies, with asymmetric
  hierarchies, wihout hierarchies - it just removes itself)
- it does not create too many problems (I am not aware of any at least)
- it does not slow down startup anymore dramatically (thanks to the fixes of
  Brian, Jeff, Tim and a significant reduction in the number of collective
  operations in the comm_query)
Any feedback is highly welcome.

This commit was SVN r7868.
2005-10-25 18:38:43 +00:00
Edgar Gabriel
00c04ab56a moving the hierarch collective component to the new parameter
registration interface.

This commit was SVN r7867.
2005-10-25 18:34:47 +00:00
Edgar Gabriel
3633605010 moving op_init further up in ompi_mpi_init, since it is required when
quering some of the collective components. Up to now, it just worked
somehow, but now with correct reference counting for ops in place, it
refused :-)

This commit was SVN r7866.
2005-10-25 18:33:48 +00:00
Jeff Squyres
ef09e768e0 Ensure to OBJ_DESTRUCT to free memory during finalize (caught by
Brian). 

This commit was SVN r7864.
2005-10-25 17:27:58 +00:00
Jeff Squyres
f8fd10715c - Minor style fix
- Be sure to properly OBJ_CONSTRUCT the intrinsic MPI_Op's
- RETAIN/RELEASE the op's when used in the invoke function

This commit was SVN r7863.
2005-10-25 16:24:00 +00:00
Jeff Squyres
a9f04c7573 Only do the extra va_* stuff if we're compiling with the compiler that
cares about it (PGI).

This commit was SVN r7860.
2005-10-25 13:08:52 +00:00
Graham Fagg
382f05c7ad Infastructure changes.
started to add static (fixed if) statement based decision rules based on gigE numbers
added mca params so that a user can force a certain algorithm/segment/topo on a per collective basis
(this is not in the fixed call path but only in the dynamic (at com create) call path).
(these params can be used by test suites such as OCC to choice which algorithm they are using).

This commit was SVN r7854.
2005-10-25 03:55:58 +00:00
Graham Fagg
d8e32464cb ops. setting/reading mca option from the right varible would help.
This commit was SVN r7850.
2005-10-24 21:33:48 +00:00
Brian Barrett
1e2f7d6a3d * make sure to expose ompi_op_t as an object
This commit was SVN r7848.
2005-10-24 20:31:14 +00:00
Rainer Keller
d6120d32d6 - Only minor white-space changes, to clean up
This commit was SVN r7843.
2005-10-24 10:36:16 +00:00
George Bosilca
b45651988b Protect against elements with ZERO length.
Remove all the useless code.

This commit was SVN r7827.
2005-10-21 06:48:51 +00:00
George Bosilca
1fb8ec646a Add the homogeneous flag back in the convertor.
Correct/improve one of the comments.
Descrease the amount of memory required for the stack.

This commit was SVN r7826.
2005-10-21 06:47:57 +00:00
Galen Shipman
cb84a57c57 add endpoint and srq flow-control..
Note, we are failing the ring tests in the intel p2p test suite, but we seem
to fail the same tests under the current trunk.. will look into this further. 

This commit was SVN r7823.
2005-10-21 02:21:45 +00:00
Galen Shipman
c4889ac759 update openib mpool to properly deregister and release (carry over from
mvapi). 

Still need to add endpoint and srq flow control as in mvapi 

This commit was SVN r7816.
2005-10-20 03:57:17 +00:00
Galen Shipman
0d1d231169 convert to new mca params, adding description strings.
changed mca param rr_buf_min/max to rd_min/max 
Add bandwidth param to openib 

This commit was SVN r7815.
2005-10-20 02:55:21 +00:00
George Bosilca
75bc3dd43c Dont mess around with the OBJ_DESTRUCT on the communicator. It's quicker (and safer) to call
directly the communicator cleanup function (ompi_convertor_cleanup).

This commit was SVN r7814.
2005-10-19 21:28:52 +00:00
George Bosilca
1d75b7972f Solve thee problem with the reference count on the datatype (RT bug 1492). The problem is that the
convertor (when prepared) increase the reference count on the used datatype. This reference count
will be released only when the OBJ_DESTRUCT is called on a convertor. However, having to call
OBJ_CONSTRUCT and OBJ_DESTRUCT on each request every time we want to use it (even when it come
from the cache) is an expensive operation. This can be avoided is the OBJ_DESTRUCT will leave the
convertor in exactly the same state as OBJ_CONSTRUCT. With this approach we just have to call
OBJ_CONSTRUCT for each convertor once when we initially create the request.

This commit was SVN r7813.
2005-10-19 20:57:39 +00:00
George Bosilca
63c5013fe6 After a OBJ_DESTRUCT a convertor has to be in a usable state. Read the comment for more informations.
This commit was SVN r7812.
2005-10-19 20:51:52 +00:00
George Bosilca
8987bcabe2 Remove the memcpy we can do it as we parse the datatypes in order to increase their references.
This commit was SVN r7811.
2005-10-19 20:51:11 +00:00
George Bosilca
6c6f17628f Remove a double OMPI_DECLSPEC from the definition of one of the predefined data-types.
This commit was SVN r7810.
2005-10-19 20:50:25 +00:00
Brian Barrett
de5e501519 Rather than hard spinning waiting for something to happen when doing shared
memory initialization, call opal_progress() to push any pending events
around and possibly yield the processor if nothing entertaining is happening.

This should probably go to the 1.0 branch.

This commit was SVN r7808.
2005-10-19 00:56:14 +00:00