Change Summary for MPICH2 for week of Aug 27, 2006

Information on all weekly change summaries is available.

List of revisions

• (penoff) implementation (prior, all channels except ch3:sctp were broke)
• (balaji) 1. MAJOR BUG -- Collective operations seem to be performing an extrapacking for derived datatypes. The first packing is for finding theactual length of the data that is being communicated, while the secondis for the actual communication. The local processor representationanyway keeps track of the actual size; we don't need to recalculateit. So, we can get rid of one of the two packings.2. MINOR BUG -- For small messages or for small communicator sizes,the collective communication directly uses send/recv operations andrelies on its packing/unpacking. However, this forces eachintermediate node to unpack the data and repack it forcommunication. Since we are anyway packing at the root node for largemessages and large communicator sizes, we can do the same for smallmessages.
• (mercierg) Added Elan and MX
• (mercierg) Added Elan and Mx stuff
• (gropp) Added Fortran mode
• (gropp) Added a timeout for the mpiexec to detect (I hope) hangs in the stdintest
• (gropp) Added a tool to print irlog files in event (really state) start time order
• (gropp) Added additional logging-debug macros
• (gropp) Added cvsignore file
• (gropp) Added earlier setting of PMI_debug and a better check for an error when reading from a new fd (as part of the init, so that failures to set up the fd will lbe detected and reported)
• (mercierg) Added elan and Mx
• (gropp) Added func enter/exit to new request and packet handler routines
• (gropp) Added generate testlist to .cvsignore
• (gropp) Added generated file
• (gropp) Added shmbase
• (gropp) Added style and copyright
• (gropp) Added two optional arg version of setfatalandjump
• (penoff) adding sctp subdir
• (mercierg) Bug fixes
• (mercierg) Changed send routine
• (gropp) Commented out the use of the non-standard directive #warning
• (gropp) Correctly determine the extension for shared libraries so that cleanlibs can find and remove them, even when a configure does not enable shared libraries
• (gropp) Dedos two files
• (mercierg) Elan vpid field added
• (gropp) Enable support for using a port in gforker mpiexec
• (gropp) Fix typo in configure test for enable-allowport
• (balaji) Fixed a a bug with MPI_Type_size() usage for heterogeneoussystems. For heterogeneous systems, MPI_Type_size() might not returnaccurate memory requirement for packing. We use MPI_Pack_size() forthis.
• (balaji) Fixed a small bug to use NMPI_Type_size().
• (gropp) Fixes to allow remshell mpiexec to specify port and to close down listener fd when all processes exit
• (robl) for some ROMIO accesses (like noncontig_coll2) we need to disable the pvfs2ncache. Certainly would be good to know what if any MPI-IO workloads benefitfrom having the ncache enabled.
• (robl) from HEAD: fixing "obviously broken" code actually introduced a bug.
• (robl) from HEAD: improved error reporting
• (robl) from HEAD: inspect file type, not memory type
• (robl) from HEAD: need to disable ncache for some workloads
• (robl) from HEAD: null out a pointer to fix a bug in EXCL case
• (mercierg) GM
• (robl) if we get an i/o error, don't just blindly stumble onwards.
• (gropp) Ignore summary.xml file in test dir
• (penoff) initial UBC SCTP CH3 commit
• (penoff) MPIDI_CH3_HAS_CHANNEL_CLOSE feature now unlocks MPIDI_CH3_Channel_close implementation (prior, all channels except ch3:sctp were broke)
• (balaji) MPI_Gather code was ignoring the existence of derived datatypes. Wewere lucky, that the MPID_Datatype_get_size_macro was returning theright value. This should be used only for contiguous datatypes.
• (gropp) Make sure that pprintirlog is executable
• (gropp) Make use of the request completion function, clean out some of the ch3 header files
• (gropp) Minor code cleanup
• (gropp) More steps toward a common set of shm basic functions for the ssm and shm channels
• (gropp) Moved unordered packet code
• (buntinas) nemesis: enable-strict code cleanup
• (balaji) Once the first pass over the datatype is done, when the next pass isbeing done (for count > 1 in the MPI_Pack call), the stored valueswere not being reset.
• (gropp) Partial clean up of shared memory code
• (mercierg) Perf improvement + bug fix
• (gropp) Removed last of the dbg_printf calls
• (mercierg) Removed unnecessary Elan libs and header
• (gropp) remshell now works as an alternate process manager
• (gropp) Restored code that was accidentally deleted in extractstates
• (mercierg) Slight changes
• (mercierg) slight changes in macros definitions
• (mercierg) Slight fixes
• (robl) sync up with head again
• (balaji) The comment in the Bcast code was inaccurate.
• (gropp) Unified the shared memory progress code from ssm and shm (files ch3_shm and shm.c respectively) into util/shmbase
• (gropp) Updated cvsignore files
• (mercierg) Various fixes
• (balaji) We can continue to use MPID_Datatype_get_size_macro. It returns the same value as MPI_Type_size() in our code.
• (balaji) When an upper layer calls Alltoallw(MPI_COMM_WORLD) with a non-zerocount for its neighbors and a zero count for the rest of theprocesses, the completion of the function should not depend onsynchronization with the remaining processes. Specifically, if theamount of data to send to (or receive from) a remote process is zero,ignore the send/recv call.

Files for each revision