This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via 384e0138cec5abac9afe9533f2d3b25a93c8eabb (commit) via 59e0d3b2971a2fcd31f82def69fa15968bbbd658 (commit) via af11500e1fbacbe31c231ee1dea8cb7bd23457b6 (commit) via b0ec78c5004601025cfb2f4d521057b80c698686 (commit) from 939601e06b888bb350e8a7bc5decc7332659ab07 (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/384e0138cec5abac9afe9533f2d3b25a93... commit 384e0138cec5abac9afe9533f2d3b25a93c8eabb Author: Pavan Balaji <[email protected]> Date: Sun May 5 21:52:44 2013 -0500 Allow devices to override the number of handles and indices through mpidpre.h. See trac ticket #1645 at mpich.org for background information, and trac ticket #1831 for discussion. https://trac.mpich.org/projects/mpich/ticket/1645 The default values for HANDLE_NUM_BLOCKS and HANDLE_NUM_INDICES in the src/include/mpihandlemem.h file can be overridden by setting the MPID_HANDLE_NUM_BLOCKS and MPID_HANDLE_NUM_INDICES #define in the adi `mpidpre.h` file. --- The motivation for this change is to allow bgq to customize how quickly the memory pools grow without changing the behavior for all other platforms. Before the changes for ticket 1645, in commits 72eb56 and 62c720, the default pool limit was 256k (256 elements * 1024 blocks) and the current default pool limit is 8192k (1024 elements * 8192 blocks). With this change the default pool limit for bgq in the pamid deivce is 2048k (256 elements * 8102 blocks). The current mpich default number of elements causes gpaw on bgq to run out of memory at scale - an increment of 1024 communicators in a pool is too large. While the new bgq defaults will limit the maximum pool size, it is unlikely that bgq will support such large pools given the bgq memory limitations. (ibm) Issue 8178 (ibm) 2ed3b95378bba964e276772bac3bb6aa45b007da diff --git a/src/include/mpihandlemem.h b/src/include/mpihandlemem.h index 584d8e8..905deec 100644 --- a/src/include/mpihandlemem.h +++ b/src/include/mpihandlemem.h @@ -97,10 +97,19 @@ const char *MPIU_Handle_get_kind_str(int kind); #define HANDLE_BLOCK_INDEX(a) ((a) & 0x00000FFF) /* Number of blocks is between 1 and 16384 */ +#if defined MPID_HANDLE_NUM_BLOCKS +#define HANDLE_NUM_BLOCKS MPID_HANDLE_NUM_BLOCKS +#else #define HANDLE_NUM_BLOCKS 8192 +#endif /* MPID_HANDLE_NUM_BLOCKS */ + /* Number of objects in a block is bewtween 1 and 4096 (each obj has an index * within its block) */ +#if defined MPID_HANDLE_NUM_INDICES +#define HANDLE_NUM_INDICES MPID_HANDLE_NUM_INDICES +#else #define HANDLE_NUM_INDICES 1024 +#endif /* MPID_HANDLE_NUM_INDICES */ /* For direct, the remainder of the handle is the index into a predefined block */ diff --git a/src/mpid/pamid/include/mpidpre.h b/src/mpid/pamid/include/mpidpre.h index 4214268..68ab472 100644 --- a/src/mpid/pamid/include/mpidpre.h +++ b/src/mpid/pamid/include/mpidpre.h @@ -66,4 +66,8 @@ #define HAVE_GPID_ROUTINES #endif +#ifdef __BGQ__ +#define MPID_HANDLE_NUM_INDICES 256 +#endif /* __BGQ__ */ + #endif http://git.mpich.org/mpich.git/commitdiff/59e0d3b2971a2fcd31f82def69fa15968b... commit 59e0d3b2971a2fcd31f82def69fa15968bbbd658 Author: Pavan Balaji <[email protected]> Date: Mon May 6 06:26:46 2013 -0500 Fix typo in hwloc's autogen.sh. No reviewer. diff --git a/src/pm/hydra/tools/topo/hwloc/hwloc/autogen.sh b/src/pm/hydra/tools/topo/hwloc/hwloc/autogen.sh index 88a122a..df42802 100755 --- a/src/pm/hydra/tools/topo/hwloc/hwloc/autogen.sh +++ b/src/pm/hydra/tools/topo/hwloc/hwloc/autogen.sh @@ -1,2 +1,2 @@ : -autoreconf ${autoreconf_args:"-ivf"} +autoreconf ${autoreconf_args:-"-ivf"} http://git.mpich.org/mpich.git/commitdiff/af11500e1fbacbe31c231ee1dea8cb7bd2... commit af11500e1fbacbe31c231ee1dea8cb7bd23457b6 Author: Pavan Balaji <[email protected]> Date: Mon May 6 06:25:41 2013 -0500 Use hwloc's autogen.sh directly. Don't just run autoreconf from hydra and use hwloc's autogen.sh, since it might do more work than just autoreconf in the future. No reviewer. diff --git a/src/pm/hydra/autogen.sh b/src/pm/hydra/autogen.sh index b84ab3b..e13fb9f 100755 --- a/src/pm/hydra/autogen.sh +++ b/src/pm/hydra/autogen.sh @@ -10,7 +10,7 @@ echo "=== running autoreconf in 'mpl' ===" (cd mpl && $autoreconf ${autoreconf_args:-"-vif"}) || exit 1 echo "=== running autoreconf in 'tools/topo/hwloc/hwloc' ===" -(cd tools/topo/hwloc/hwloc && $autoreconf ${autoreconf_args:-"-vif"}) || exit 1 +(cd tools/topo/hwloc/hwloc && ./autogen.sh) || exit 1 echo "=== running autoreconf in '.' ===" $autoreconf ${autoreconf_args:-"-vif"} || exit 1 http://git.mpich.org/mpich.git/commitdiff/b0ec78c5004601025cfb2f4d521057b80c... commit b0ec78c5004601025cfb2f4d521057b80c698686 Author: Pavan Balaji <[email protected]> Date: Sun May 5 20:46:39 2013 -0500 Reorder compiler preference. clang seems to have the same or slightly better performance compared to gcc, which, in turn has the same or slightly better performance compared to cc. No reviewer. diff --git a/confdb/aclocal_cc.m4 b/confdb/aclocal_cc.m4 index 98fdc34..f8ace03 100644 --- a/confdb/aclocal_cc.m4 +++ b/confdb/aclocal_cc.m4 @@ -11,7 +11,7 @@ AC_DEFUN([PAC_PROG_CC],[ dnl developers notice this case. AC_BEFORE([$0],[AC_PROG_CC]) PAC_PUSH_FLAG([CFLAGS]) - AC_PROG_CC([icc pgcc xlc xlC pathcc cc gcc clang]) + AC_PROG_CC([icc pgcc xlc xlC pathcc clang gcc cc]) PAC_POP_FLAG([CFLAGS]) ]) dnl ----------------------------------------------------------------------- Summary of changes: confdb/aclocal_cc.m4 | 2 +- src/include/mpihandlemem.h | 9 +++++++++ src/mpid/pamid/include/mpidpre.h | 4 ++++ src/pm/hydra/autogen.sh | 2 +- src/pm/hydra/tools/topo/hwloc/hwloc/autogen.sh | 2 +- 5 files changed, 16 insertions(+), 3 deletions(-) hooks/post-receive -- MPICH primary repository