Commits · c655784c9703992d037562fac82ad0536b3e9fe1 · Rob Latham / MPICH-BlueGene

18 May, 2013 9 commits

White space cleanup. · c655784c
Pavan Balaji authored May 18, 2013
```
No reviewer.
```
c655784c

Added new string stash/spit macros. · 6a39ab87

Pavan Balaji authored May 02, 2013

We have been allocating a "large enough" pool of temporary strings and
hoping that we don't overflow it for our temporary allocations.  This
commit provides macros that better check for overflow problems without
making the code too clumsy.

No reviewer.

6a39ab87

test: add attrdeleteget, MPI_Attr_get called from delete_fn · c532aacc

Jed Brown authored May 18, 2013 and

Pavan Balaji committed May 18, 2013



With MS-MPI 64-bit from HPC Pack 2008 and 2012, MPI_Attr_get returns
error code 773 when called from delete_fn on a communicator obtained
from MPI_Comm_split.

The standard is not explicit that the 'comm' argument of delete_fn must
be valid, so this test is only in effect when !USE_STRICT_MPI.

Signed-off-by: Jed Brown <jedbrown@mcs.anl.gov>
Signed-off-by: Satish Balay <balay@mcs.anl.gov>

c532aacc

config: change PAC_FUNC_NEEDS_DECL to check by assignment · 76b46a77

Jed Brown authored May 17, 2013 and

Pavan Balaji committed May 18, 2013



The old method attempts to create a failed compile when the function
*is* already declared, by creating an incompatible declaration.  That is
insufficient with clang-3.2, which only warns in case of incompatible
redeclaration of a library function:

  conftest.c:44:5: warning: incompatible redeclaration of library function 'strncmp'
  int strncmp(double, int, double, const char *);
      ^
  /usr/include/string.h:143:12: note: 'strncmp' is a builtin with type 'int (const char *, const char *, size_t)'
  extern int strncmp (const char *__s1, const char *__s2, size_t __n)
             ^
  1 warning generated.

Instead, we take the opposite approach, generating a failed compile when
the library function has *not* been declared in the header.

  void (*fptr)(void) = (void(*)(void))strncmp;

Signed-off-by: Jed Brown <jedbrown@mcs.anl.gov>

76b46a77

Hardcoded to wrong value 400000 if MP_POLLING_INTERVAL not set · 533b8b37

Su Huang authored Apr 17, 2013 and

Pavan Balaji committed May 18, 2013

(ibm) D189371
(ibm) b02a38e0680dbd1471884439214ee6afe13baeb4
(ibm) c8ebe0da80732297de985c3640caa4c5623c55d4
(ibm) 48d1993041c3820a468d2ee6c838ff0a5bc1154a

533b8b37

pami_tune -o option renames directory · 11fecdc3

Sameh Sharkawi authored Apr 22, 2013 and

Pavan Balaji committed May 18, 2013



 1 - Fixes for command line paramters
 2 - Fix for allgather PAMI_IN_PLACE handling in collective selection path

(ibm) D189995
(ibm) 7Z8
(ibm) bad205a6147cebce9a40f3b50560df80d3ddb354

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

11fecdc3

One more fix for correctly displaying warnings in tune mode · ac596401

Sameh Sharkawi authored Apr 18, 2013 and

Pavan Balaji committed May 18, 2013


+ build fix

(ibm) D189887
(ibm) 556b0b576c22ef4cc8b745d2ebc57bf4cf354818
(ibm) cee6d5876ad57094c8631e6c3faaef1bb213377a

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

ac596401

Changing default iter count to 1000 · 8ac56947

Sameh Sharkawi authored Apr 18, 2013 and

Pavan Balaji committed May 18, 2013



(ibm) D189889
(ibm) 7Z8
(ibm) b987ee6281ea76879793f596a0f6c93c425716d7

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

8ac56947

Alltoallv_int is missing in the xml file · 02b0a83f

Sameh Sharkawi authored Apr 18, 2013 and

Pavan Balaji committed May 18, 2013



  1 - A fix where all (v) algorithms are disabled in collective selection
      and  only (v)_int algorithms are enabled since MPICH only supports
      (v)_int algorithms. IN the tune case, (v) collectives are skipped
      even if user selected them. Also a warning message is displayed to stderr
      if infolevel is >= 1 in collective selection and tuning cases.
  2 - Another fix for one task geometries where the fast_query object was
      not initialized but was still used and may cause core dump.

(ibm) D189887
(ibm) 7Z8
(ibm) 0ea3f16a9fd6e2fbc1cfffeb5401cce20a861e96

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

This is the 2nd commit message:
  BGQ build break fix

02b0a83f

17 May, 2013 8 commits
- Allow pamid to override max message size for SMP collectives. · cf31e37a
  Pavan Balaji authored May 15, 2013
```
Reviewed by Charles Archer @ IBM.
```
  cf31e37a
- Enable SMP aware collectives via PAMI local task detection · cbcde2f4
  Charles Archer authored Mar 11, 2013 and Pavan Balaji committed May 17, 2013
```
(ibm) D188808
(ibm) 2021904184004ba65cb4f277d01dc88b47d692bc

Signed-off-by: Bob Cernohous <bobc@us.ibm.com>

Modified significantly by Pavan Balaji to utilize mpich infrastructure
more consistently for all SMP collectives.

Reviewed by Charles Archer @ IBM.
```
  cbcde2f4
- Cap the maximum SMP collective size. · e279b1ae
  Pavan Balaji authored May 15, 2013
```
Reviewed by Charles Archer @ IBM.
```
  e279b1ae
- Added ability to enable/disable individual SMP collectives. · 60a89bbd
  Pavan Balaji authored May 15, 2013
```
Reviewed by Charles Archer @ IBM.
```
  60a89bbd
- Make enabling SMP collectives an environment variable instead of a · 4456bbec
  Pavan Balaji authored May 15, 2013
```
configure option.

Reviewed by Charles Archer @ IBM.
```
  4456bbec
- Update parameter generation to allow device-overrides where needed. · 430514fd
  Pavan Balaji authored May 15, 2013
```
Reviewed by Charles Archer @ IBM.
```
  430514fd
- Better arrange environment variables according to their categories. · c6084635
  Pavan Balaji authored May 15, 2013
```
Reviewed by Charles Archer @ IBM.
```
  c6084635
- Make parameter names consistent. · fa950cb5
  Pavan Balaji authored May 14, 2013
```
All CH3 parameters start with CH3_ now.  All nemesis parameters start
with NEMESIS_.  For netmod specific parameters, we use
NEMESIS_<netmod>_.

Reviewed by Charles Archer @ IBM.
```
  fa950cb5
16 May, 2013 1 commit
- Reverted [c478a4fe] which seems to be causing build failures. · 50d4516d
  Pavan Balaji authored May 16, 2013
```
No reviewer.
```
  50d4516d
15 May, 2013 17 commits

XL compiler prototype fix · 1288cced
Charles Archer authored May 14, 2013 and Pavan Balaji committed May 15, 2013

1288cced

updates for token flow control · faf3fcc0

Su Huang authored May 01, 2013 and

Pavan Balaji committed May 15, 2013



(ibm) D190238
(ibm) 1fe9a013d48ef38d1505dd96977fbe105b2a2b90

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

This is the 2nd commit message:
  BGQ build break fix

faf3fcc0

32bit(2) reduce&allreduce fail MPI_LONG w/ collective_selected · c756a7d8
Su Huang authored Apr 29, 2013 and Pavan Balaji committed May 15, 2013
```
(ibm) D189842
(ibm) 7Z8
(ibm) b4718cf2b8bf3ddb7d39e2fdf7aee7b87eac8e1f

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>
```
c756a7d8

32bit reduce&allreduce fail MPI_LONG w/ collective_selected · afa9629e

Su Huang authored Apr 29, 2013 and

Pavan Balaji committed May 15, 2013



(ibm) D189842
(ibm) 7Z8
(ibm) 017c9d99634bcfb225c4ede785709e3f3644a193
(ibm) a42fc06e0ed99acd968fc18405199bc98e5b1190
(ibm) 5326308445643a91cdd15ec7b402285aef9095e4

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

afa9629e

second spawn reuse taskid would fail with ERROR 421 · 93580586
Haizhu Liu authored Apr 22, 2013 and Pavan Balaji committed May 15, 2013
```
(ibm) D189968
(ibm) 4260cf9294bc75af0cb175823885f20d1d2b1ee8

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>
```
93580586
Correctly handling non-continuous data in vector collectives · 978d1c70
Sameh Sharkawi authored Apr 15, 2013 and Pavan Balaji committed May 15, 2013
```
(ibm) D189791
(ibm) 921af973d312be39763878d9478bea10187ad762

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>
```
978d1c70
MPICH missing timer only progress engine that old PEMPI has · 7da2069f
Su Huang authored Apr 16, 2013 and Pavan Balaji committed May 15, 2013
```
(ibm) D189783
(ibm) 769e26acc31c01fdda183fe86f871df73504064c

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>
```
7da2069f

pami_tune command line changes · 1820f7c5

Sameh Sharkawi authored Apr 15, 2013 and

Pavan Balaji committed May 15, 2013



    * pami_tune's -v  flag conflicts with poe's -v flag.
    * Disable -x flag for checkpoint mode until later

(ibm) D189881
(ibm) 749763c4d0c973739c356d59d561df6feb40aad3

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

1820f7c5

Check for assert only if asserts are enabled. · bfa19b44
Michael Blocksome authored May 08, 2013 and Pavan Balaji committed May 15, 2013

bfa19b44
Use pami_endpoint_t for root instead of incorrect pami_task_t · 5a63aaf0
Charles Archer authored Apr 14, 2013 and Pavan Balaji committed May 15, 2013
```
(ibm) D189828
(ibm) 51d29ce7faf74d2d2c8a02798cd7880e1a5f3783

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>
```
5a63aaf0

Need service variables for all collectives · d92aabf2

Sameh Sharkawi authored Apr 12, 2013 and

Pavan Balaji committed May 15, 2013



(ibm) D189833
(ibm) 7Z8
(ibm) 9e89a54cd839c58e0c0573b09f39e4adf6e012e3

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

d92aabf2

Thread level promoting w/dyn_tasking due to world exiting handler in separate thread · 662d6ff0
Haizhu Liu authored Apr 12, 2013 and Pavan Balaji committed May 15, 2013
```
(ibm) D189766
(ibm) 623ef12fe24e80752c51b38db46926520e530136

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>
```
662d6ff0
duped info in wrong branch of check for null · 63e1d88f
Rob Latham authored May 15, 2013

63e1d88f
deal with MPI_INFO_NULL reasonably · 1b5334ed
Rob Latham authored May 15, 2013
```
maddeningly we cannot dup MPI_INFO_NULL, so let's check for that case
```
1b5334ed
Remove C99-ism from IBM contributed code. · 49595414
Pavan Balaji authored May 14, 2013
```
Reviewed by Mike Blocksome @ IBM.
```
49595414

Naming updates to IBM's NBC patch. · 57f5b7b1

Pavan Balaji authored May 14, 2013

We now provide two types of function pointers: _sched and _req.  The
_sched functions just add instructions to the schedule, but no request
is created.  The _req functions return an MPID request.

Reviewed by Mike Blocksome @ IBM.

57f5b7b1

Add hooks for optimized MPIX_* non-blocking collectives. · 2eefd3a2

Michael Blocksome authored Nov 01, 2012 and

Pavan Balaji committed May 15, 2013



The following collectives are updated:

  MPIX_Ibcast
  MPIX_Iallgather
  MPIX_Iallgatherv
  MPIX_Iallreduce
  MPIX_Ialltoall
  MPIX_Ialltoallv
  MPIX_Ialltoallw
  MPIX_Iexscan
  MPIX_Igather
  MPIX_Igatherv
  MPIX_Ireduce_scatter_block
  MPIX_Ireduce_scatter
  MPIX_Ireduce
  MPIX_Iscan
  MPIX_Iscatter
  MPIX_Iscatterv

These collectives are updated similar to how the optimized MPIX_Ibarrier
is implemented:

| The previous MPIR_Ibarrier_impl() function forced all adi implementations
| to create a MPID_Sched_t opaque object which was then passed in to the
| specific ibarrier implementation via a function pointer table.
|
| The MPID_Sched_t object represents a completely new state machine that
| must be advanced whenever mpi progress is made.
|
| The required construction of the MPID_Sched_t object and the required
| advance of the schedule state machine would be extremely detrimental to
| pamid performance.

(ibm) 47b497e15cd87df666c6031a28f5f8a5ae53cece
(ibm) aee872eb91f4300066957349e5a3d8103da33f40

Signed-off-by: Charles Archer <archerc@us.ibm.com>

2eefd3a2

14 May, 2013 4 commits

Warning squash. · 635969e7
Pavan Balaji authored May 14, 2013
```
No reviewer.
```
635969e7

Disable -Wcast-align and -Wshorten-64-to-32. · c537d620

Pavan Balaji authored May 14, 2013

This is an important warning flag, but given the number of places we
are breaking this in mpich, it is causing us to miss other warnings.
This is meant to be a temporary disabling.

No reviewer.

c537d620

adio/common wrong checking of lseek return · 74639d55
Haizhu Liu authored May 01, 2013 and Rob Latham committed May 14, 2013
```
(ibm) D190161
(ibm) 55566af07e7d69df17005e50c783ff3ae26cc325

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>
```
74639d55

avoid hang if subset of procs have invalid INFO · 470667cb

Rob Latham authored May 14, 2013

In response to IBM's integration ticket 1822 (but reworked), turn
MPIO_CHECK_INFO into a collective macro; exchange result of info
inspection with all procs.  Now a bogus info on one proc won't cause a
process hang.

470667cb

13 May, 2013 1 commit

multi_mpi coredump with RDMA on mpich · 49195779

Su Huang authored Sep 21, 2012 and

Pavan Balaji committed May 13, 2013



(ibm) D186334
(ibm) 7Z6
(ibm) 63716b02ccab7cd99116e927a210eca43b03d52e

Signed-off-by: Michael Blocksome <blocksom@us.ibm.com>

49195779