Change Summary for MPICH2 for week of Aug 13, 2006

Information on all weekly change summaries is available.

List of revisions

• (balaji) 1. Fixed the transp-datatype.c matrix transpose test program toactually perform the transpose. The example, was sending and receivinginto the same datatype, resulting in the matrix being transferred asit is, instead of in the transposed format.2. The datatype packing scheme allocates 256KB of memory chunks forevery send, wasting the extra space for smaller sends. Fixed this toonly allocate as much as needed.
• (balaji) actually perform the transpose. The example, was sending and receivinginto the same datatype, resulting in the matrix being transferred asit is, instead of in the transposed format.2. The datatype packing scheme allocates 256KB of memory chunks forevery send, wasting the extra space for smaller sends. Fixed this toonly allocate as much as needed.
• (gropp) Add initialization of MPI objects when memory tracing is enabled
• (gropp) Add support for generating asm output to use in performance tuning
• (chan) added -dmode=null in install target
• (gropp) Added MPI_Wtime/Wtick to declarations
• (gropp) Added a new internal error when initializing the array of packet handlers
• (gropp) Added definition to the conf file so that includes can check whether the conf file has been included before making use of the values defined in the conf file
• (gropp) Added ifdefs for MPID_HAS_HETER, then disabled them to support external 32, plus added missing hetero_dloop initialization to MPID_Type_dup. Left in the ifdef MPID_HAS_HETERO in case it may be useful when completing heterogeneous support
• (gropp) Added implicit none to test for definition of wtime and pmpi_wtime (and wtick)
• (gropp) Added info on Intel v9 ifort
• (thakur) added lookup_name test
• (chan) Added support so that MODE and XMODE can be empty. Added DMODE as permission for directory and -dmode= as input argument to the script
• (gropp) Added testing targets
• (thakur) added unpublish test
• (gropp) Added verbose mode output
• (penoff) adding Nemesis SCTP net_mod
• (penoff) adding sendQ's to Nemesis sctp net_mod
• (balaji) allocating a buffer each time.
• (thakur) applied 1.0.4p1 updated here (implicit none)
• (gropp) Applied fix from 1.0.4 p1 for declaring wtime/wtick in Fortran 90
• (mercierg) Assert added
• (gropp) Clean up debug messages
• (mercierg) Current working version
• (mercierg) Elan stuff
• (gropp) Fix typo in comment
• (balaji) Fixed a bug in the dataloops code for datatype communication. The code was first creating an IOV list for the first 16 elements and checking its density. If the density is low, it packs the data. Now, both the creation of the IOV and the packing of the data go through the same segment manipulation code. So, the creation of the IOV modifies the state of the stack (i.e., till what point in the datatype it has proceeded). Later when the data packing is done, the previous segment manipulation (done for the IOV creation) is ignored -- so the code was getting confused and going over the entire datatype again. Fixed this to copy the data in the IOVs separately, so that we start from whichever point in the datatype the IOV creation stopped.
• (balaji) Fixed a bug in the memory allocation patch.
• (balaji) Fixed a small initialization bug.
• (thakur) fixed bug in Makefile that caused bibtex to fail in user guide
• (rbutler) Fixed misspelling of rhsSocket to be rhsSock
• (gropp) Fixed pingping to limit total message size and added a test that Isend does not block waiting for the matching receive
• (gropp) Improved error message
• (gropp) Include the mpidi_ch3_impl.h file first
• (gropp) Initialize name and error handler fields
• (gropp) Make the MPD (really PMI) version of MPID_NS_Create return a valid handle on success
• (chan) misc cleanup and removed the beginning clean target in the default build.
• (gropp) Move RMA structures into a separate include file and add discussion about streamlining request management
• (gropp) Move call to obj_alloc_complete out of the critical path on object allocation
• (thakur) moved spawn error tests to errors/spawn directory
• (gropp) Name this the 1.0.4p1 version
• (buntinas) nemesis: fixed error messages
• (buntinas) nemesis: more new tcp stuff
• (buntinas) nemesis: newtcp updates
• (mercierg) New send
• (gropp) Remove dead ch3_request functions and update common ch3u_request.c
• (gropp) Removed C++ style comments and changed asm volatile to __asm__ __volatile__ to match the test for __asm__ __volatile__ in the nemesis configure (also required for nemesis to compile when --enable-strict selected)
• (gropp) Removed MPIDI_CH3_Request_create in favor or MPID_Request_create, as some CH3 code used one and some used the other, and the inconsistent use caused confusion. This also takes us one step closer to optimized request creation (initializing only the necessary fields)
• (chan) removed beginning clean_lib and clean_bin(MPICHx specific) targets in the default target
• (chan) removed the beginning clean target in the default build.
• (mercierg) Small fix
• (chan) split object files into 2 different types: .p@OBJEXT@ and .n@OBJEXT@ where former refers to parallel(mpi) object and later refers to normal(serial) object. Reworked all library targets so object files, executables and libraries won't be recreated in consecutive makes.
• (thakur) updated
• (gropp) Use a leaner request initialization
• (chan) used MV set by configure in impliciit riles
• (chan) Used autoconf marcos @OBJEXT@ and @EXEEXT@. Replaced static target names by the corresponding library and executable pathname so the library and executables won't be recreated in consecutive makes.
• (balaji) Using a free buffer pool for packing non-contiguous buffers instead ofallocating a buffer each time.
• (mercierg) Working version

Files for each revision