The MPI/GAMMA home page |
Minor changes needed to compile with latest release of GAMMA.
Those changes makes now possible to submit MPI/GAMMA jobs through
Torque/PBS via a "mpiexec" utility.
For this you need a patched version of the Torque/PBS "mpiexec"
utility (derived from mpiexec version 0.83), whose source code can be
downloaded
here.
Minor fixes needed to compile with latest release of GAMMA.
Application-level performance comparisons (always reported under Section Performance in the GAMMA page): OpenFOAM 1.4 (more to follow)
Minor fix for compiling with gfortran.
Fixed a couple of name clashes wrt the ROMIO internal layer for parallel I/O.
Improved MPI_Alltoall(), thanks to Tony Ladd (Univ. Florida).
Fixed a bug of rendez-vous communications in the case of self-communication in a subgroup.
Fixed another bug in the management of per-message credits.
This version of MPI-GAMMA is based on MPICH 1.2.7, subversion ``p1''. After years with the old (and a bit buggy) MPICH 1.1.2, I decided to move to the last MPI-1 release of MPICH. Hopefully this will bring more stability and completeness at the MPI level.
Fixed a bug in the management of per-message credits. This caused some programs to abort with a "out of memory" error (actually it was sort of "out of credits").
Using dynamically allocated memory for unexpected messages, instead of the old static memory pool.
Shared versions of the MPI libraries (both C and F77 bindigs) are built, in addition to the static ones.
Some changes to the "configure" script. Please read CAREFULLY the README.GAMMA file inside the source package, once expanded onto the MPICH tree.
Found and fixed a bug into src/util/bsendutil2.c (buffer management inside MPI_Bsend()).
Internal flow control algorithm has been improved.
MPI Abort routine now really aborts.
Various bugs have emerged while porting the system to 64-bit architecture. This version of MPI/GAMMA is thus able to work with GAMMA on x86_64 as well as the good old 32-bit :-)
This version is for the new release of GAMMA, in which the API has been slightly changed.
Added more info to the README.GAMMA file.
Minor changes to avoid annoying compiler warnings.
Main Makefile slightly changed.
The use of GAMMA native bcast() routine has been disabled again (see file mpich/src/coll/intra_fns.c.gamma), as it seems to be not that efficient anyway. So, the current version of MPI/GAMMA is equivalent to the one dated 17 July 2002.
The use of GAMMA native bcast() routine is now enabled by default (see file mpich/src/coll/intra_fns.c.gamma), thanks to improved stability of GAMMA flow control with broadcast messages.
The size of the eager-to-rendezvous threshold has been increased to 256 KByte.
The use of GAMMA native bcast() routine is still disabled by default (see file mpich/src/coll/intra_fns.c.gamma).
Minor changes, due to minor changes in the GAMMA API.
The size of some internal queues has been increased.
The use of GAMMA native bcast() routine is still disabled by default (see file mpich/src/coll/intra_fns.c.gamma).
The slightly more aggressive gcc optimization options in Makefile have been turned off (the compiler egcs-2.91.66 apparently produces wrong code when the option -mpentium is used).
The use of GAMMA native bcast() routine is still disabled by default (see file mpich/src/coll/intra_fns.c.gamma).
Slightly more aggressive gcc optimization options in Makefile.
The use of GAMMA native bcast() routine is still disabled by default (see file mpich/src/coll/intra_fns.c.gamma).
The use of GAMMA native bcast() routine has been disabled by default, because it seems not stable enough. See file mpich/src/coll/intra_fns.c.gamma
The use of GAMMA non-blocking send routines has been disabled by default, because the new version of GAMMA provides retransmission of missing packets but only with blocking communication. Non-blocking routines of GAMMA are thus less reliable than blocking ones; it is however possible to use them with MPI/GAMMA, by acting upon constant USE_GAMMA_ISEND in file gamma_def.h (in the source code distribution of MPI/GAMMA).
Bug fix in the memory lock/unlock code.
Minor code restyling.
The bug that was used to show up with MPI/GAMMA numerical jobs when running with IRQ-driven receive enabled has been discovered to be a bug of GAMMA (and not of MPI/GAMMA, as I wrongly was believed it was).
Found and fixed a bug in the rendez-vous communication protocol of MPI/GAMMA ("send-to-self" case).
Minor bug fixes.
Fixed a bug in the "mpigammarun" script for job launch (to be used in place of the standard "mpirun").
Minor improvement over the source distribution.
Modification to the "mpigammarun" script for job launch (to be used in place of the standard "mpirun").
This version uses the GAMMA non-blocking send() routines with the MPI non-blocking communication calls, wherever feasible.
Added missing information to the README.GAMMA file (inside the source code distribution) concerning the script "mpigammarun".
Minor changes.
Found and fixed a bug in function MPID_Iprobe() (file adi2probe.c), which broke the MPI_Iprobe() routine (a necessary device polling cycle was missing).