Change Summary for MPICH2 for week of Jun 24, 2007

This list was automatically generated on Mon Jun 25 04:10:29 2007 .

Information on all weekly change summaries is available.

List of revisions

Files for each revision

added --enable-safePMPI support in threaded MPI logging through the thread-specific MPE logging on/off switch. Moved MPE_LOG_ON/MPE_LOG_OFF outside of global mutex in all MPI functions being profiled. src/mpe2/src/wrappers/src/log_mpi_core.c src/mpe2/src/wrappers/src/log_mpi_io.c src/mpe2/src/wrappers/src/log_mpi_rma.c src/mpe2/src/wrappers/src/log_mpi_spawn.c
Added a test for MPI_SIZEOF test/mpi/f90/datatype/Makefile.ap test/mpi/f90/datatype/indtype.f90 test/mpi/f90/datatype/sizeof.f90 test/mpi/f90/datatype/testlist.ap
Added a test to reproduce the performance problem with performing gather with a non-zero root. test/mpi/perf/non_zero_root.c test/mpi/perf/testlist
Added a thread-specific MPE logging on/off switch and its related support. This should enable thread logging with --enable-safePMPI in MPI logging. src/mpe2/src/logging/include/mpe_log_thread.h.in src/mpe2/src/logging/src/mpe_log_thread.c
Changed recv_dict_msg to loop around receiving 8-byte len. src/pm/mpd/mpdlib.py
changed the default --enable-threadlogging=yes. src/mpe2/src/logging/configure.in
Changed the error margin to 1.0 instead of 0.5. This is pretty much a random change to eliminate false positives. We need a better way of coming up with this number, or a better test itself. test/mpi/perf/sendrecvl.c
Define *_FILENO on windows -- defined in <unistd.h> on unix src/mpe2/src/logging/include/clog_const.h
Don't have to deal with single element vectors separately. The loop unrolling code handles it. src/mpid/common/datatype/dataloop/segment_ops.c
Eliminate duplicate IPs for a single host from gethostbyname_ex; keep unique ones. src/pm/mpd/mpdboot.py
Fixed a bug in the calculation of the buffer size for gather. src/mpi/coll/gather.c
Fixed a bug in the loop unrolling code. src/mpid/common/datatype/dataloop/segment_ops.c
fixed comment src/mpe2/src/wrappers/configure.in
For "historical" reasons, define WIN32_MEAN_AND_LEAN before including windows.h if you plan to include winsock2.h later --- to prevent inclusion of winsock.h from within windows.h src/util/thread/win_types.i
Make sure that there the results are not too noisy. Also, run the benchmark multiple times and take its average. test/mpi/perf/dtpack.c
Minor fix for f90 integer type determination src/binding/f90/configure.in
Modified the benchmark to allow for 1-2 errors without reporting them. This is a hack to stop false positives from cache miss effects, eager-rendezvous threshold jump, etc. test/mpi/perf/sendrecvl.c
Moved the createf90types test to a different directory test/mpi/f90/datatype/testlist.ap
Moved the details of veccpy into the primary macro instead of a second redirect (which was not doing anything additional). src/mpid/common/datatype/dataloop/segment_ops.c
nemesis: fixed order of 63/32-bit library paths src/mpid/ch3/channels/nemesis/configure.in
removed the code checked in revision 1.40. i.e. restored back to the case that when --enable-threadlogging=yes, then --enable-safePMPI=yes for non-MPICH2 build. This is made possible because of the thread-specific MPE logging switch. src/mpe2/configure.in
Setting socket state to only readable during the RANKSENT stage. src/mpid/ch3/channels/nemesis/nemesis/net_mod/newtcp_module/socksm.c src/mpid/ch3/channels/nemesis/nemesis/net_mod/newtcp_module/socksm.c
updated src/mpi/romio/common/dataloop/Makefile.in
Use proper type for address ints test/mpi/f90/datatype/createf90.f90
Workaround for a bug in some versions of autoconf that discard trailing blanks in statements - this broke code that change the IFS confdb/aclocal_f77.m4