Change Summary for Revisions between MPICH2_1_0_5 and MPICH2_1_0_6

List of revisions

• to compile and run CPI
• (1) Added check for the latest version of VS runtime dlls (2) Updated the link for obtaining the .NET framework -- now all http links in installables are generic
• (1) Corrected the estimation of the temporary buffer size to be more accurate. (2) Cut down a memory copy operation on the pseudo-leaf nodes for communicators which are not a power of 2.
• (apear to) clobber extra_state for some mysterious reason with certain C++compilers.
• -- Minor modifications to allow IB net module to compile and run CPI
• -- Minor modifications to allow IB net module to compile and run CPI
• 1) Added _FILENO* in mpe_logging_conf.h for x64 (2) Added for include path related to MPE -- added MPE callstack related include paths -- for x64 makefiles (3) Added MPE callstack related src files to x64 makefiles
• 1) Added linker flag /MANIFEST to generate manifest files in GenBuild() 2) Added target to create fmpe directory if it does not exist - in GenBuild()
• 1) Adding mkstemp for windows (2) Retaining the file stream
• 1) Changed path of shmproc.c from ch3/util/shm to ch3/util/shmproc 2) Removed ssm/src/ch3i_get_business_card.c -- syncing with changes in source tree structure
• 1) Changes related to generalized requests - Async IO on windows (2) Thread safe IO error messages on windows
• 1) Corrected macro definition indicating inclusion of a header file - WAS - #define *CONF_H - IS NOW -#define *CONF_H_INCLUDED 2) Creating ch3usockconf.h
• 1) Create mpe_callstack.h 2) Handle @mpe_callstack*@ --- MPE Callstack impl is NOT complete on windows
• 1) Create mpich2info.h and handle template strings for configure, compiler arguments (2) Add src dir for shm channel (3) Modified CreateObjs() to handle simple makefiles of shm & sshm channels correctly
• 1) Removed ib & sshm channels (2) Removed/Updated configurations referring ib & sshm channels (3) Corrected the Ctrl chars on makewindist.bat -- now compatible with windows (4) Now the default build is the Release build
• 1) Spawned processes inherit working directory from spawner if the spawner does not specify it explicitly in MPI_Info
• A stub for NTFS async I/O
• AC_DEFINE( HAVE_LIBPTHREAD ) since AC_CHECK_LIB is used with "action-if-found", i.e. AC_CHECK_LIB won't define HAVE_LIBPTHREAD as usual.
• Add -libtype optiona and more tests for valid major and minor versions
• add a prototype for MPIR_Grequest_progress_poke
• Add code to try and determine the Fortran name mangling when C cannot be used as the linking language (e.g., when using the PG Fortran 90 compiler as Fortran)
• add explicit rule for wrong_int_byte to avoid compilation by serial C compiler.
• add in code to deal with NFS brokenness
• Add more detailed debugging message for unexpected incomplete send on close protocol
• Add testing target to f77 Makefile.sm
• add the mpix_grequest_class_create error message
• Add the needed include files for prototypes
• Add try block around send_dict_msg
• add two aggregation tests, exercising the "force cb_writes" path
• Add undefine of MPIU_CALL for internal references in the dynamically loaded channel
• Added "-mapall" option to mpiexec to map all network drives currently mapped by the user executing mpiexec command
• Added "sleep 1" between making each subpackage to avoid the time resolution problem in Makefile.
• added --enable-coarsethread to forcefully turn on/off of the coarse-grained thread support
• 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.
• added --with-cross=file support.
• added -lm
• added -lpthread when coarse-grained pthread logging support is enabled for non-MPICH2. Added configuration file mpe_log_postlib.conf to append -lpthread and enable mpe_mpilog.conf to read from mpe_log_postlib.conf.
• Added ABI Version information to channels
• Added BIGMEM version of the allocation macros along with an --enable-alloca to enable the use of alloca for small-memory allocation
• Added C tests of the type_create_f90 functions
• Added CH3_MSG for debugging message headers and envelops; also fixed the creation of filenames for the mpich-dbg options
• Added CPPFLAGS to the targets that produce the preprocessed source files
• Added CallStack subpackage support.
• Added ERR_SPAWN to known errors list
• added Fortran interface to MPI_Init_thread()
• added Fred Webb's TotalView related comments
• Added GCC_OPTFLAG to control gcc optimization, particularly with --enable-strict
• Added InitCompleted, VC_Destroy, and PG_Destroy functions, including to dll functions
• Added MPE callstack -- default dummy impl -- related src files and modified include path
• added MPE changes
• added MPE logging for ADIOI_WRITE_LOCK, ADIOI_READ_LOCK and ADIO_LOCK and a special "postwrite" state for synchronization in MPI-IO collective call.
• added MPE logging for pvfs2's open (i.e. lookup/create).
• added MPE logging instrmentation for PVFS_sys_read() and PVFS_sys_write().
• Added MPE logging instrumentation to the IO calls, open/read/write/lseek/close,used in the NFS code.
• added MPE related documentation.
• added MPE_Initialized_logging to keep NEC SX-8's fortran compiler happy.
• added MPE_KIND_MSG_INIT into allowed_mask, so MPI_Send_init()/MPI_Recv_init + MPI_Startall() will work on persistent requests
• Added MPICH2 version define
• Added MPICH2_INCLUDE_FLAGS to CPPFLAGS in smpd configure
• Added MPI_OFFSET_KIND_VAL to ensure that a valid kind parameter is used in tests of Fortran (before, in some cases an invalid Fortran program was generated with integer (kind=), that is, with an empty string for the kind value)
• Added SHM fixes with the memory barriers and volatile structure elements to the SSM channel too.
• added a HEADER argument to PAC_MPI_LINK_CC_FUNC
• Added a combined send/recv poll function. The current version is not very efficient since the poll always returns immediately saying that there is available space to write, irrespective of whether there is data to write or not.
• Added a fix for the Fortran 90 tests for external routines
• Added a fixme
• Added a fixme about broken intermodule interface
• Added a fixme question
• Added a hook to allow the device to complete any initialization
• Added a manual test for a good Dims_create
• Added a mechanism so that MPE LDFLAGS, -Wl,--export-dyanmic, can be ignored in using MPE wrappers. Updated the documentation as well.
• Added a mention for the memory leaks.
• Added a mpich2version program
• Added a nesting-debug check for negative nest count
• Added a new MPID function
• Added a new test to check and set the address-sized integer to one of the CLOG_intXX_t and use it in mpi_null.h to avoid 64-bit gcc from complaining.
• added a new test, log_cost.c, which is linked by libmpe_nompi.a to test the serial MPI implementation. The log_cost.slog2 also shows the overhead of MPE logging.
• Added a numeric MPICH2 version
• Added a prototype for mkstemp to keep ANSI C happy.
• Added a send-flood test, particularly for shared memory methods
• added a separate section on MPE, documented the compiler wrapper's linking switches.
• Added a step toward support for intercomm split
• Added a test for Ibsend and Request_free
• Added a test for MPI_SIZEOF
• Added a test for error checking on dims size in cart create
• Added a test for no fildes available; cleaned up some use of err and errno, added handling for EINTR in one place, added a temporary FIXME fprintf for a case that will cause problems but may be unlikely
• Added a test for orphaned executables
• Added a test for splitting an intercommunicator (allowed in MPI-2)
• Added a test for truncated short messages
• Added a test of type dup of a derived datatype including an array of types
• Added a test that the autoconf macro correctly determined the FLIBS values
• Added a test that user-defined error codes are correctly returned when used in an attribute copy routine
• added a test to check gethostname() declaration
• Added a test to help find memory leaks in type_dup
• Added a test to reproduce the performance problem with performing gather with a non-zero root.
• Added a union to ensure alignment of data payload.
• Added a version of the MPICH1 sendmany test; this found a bug in the eager short handling
• Added a way to import definitions from the channel for just the ch3 device, plus some cleanup
• Added additional check to connection private data to avoid SEGV during disconnect
• Added an MPIU_CALL to allow either static or dynamically loaded library calls
• Added an approximate block for publishing the process pid and hostname in the kvs space, to allow other processes to access it and build a table to help the debugger
• added an empty HEADER argument to match the modified PAC_MPI_LINK_CC_FUNC
• Added an entry for the non_zero_root benchmark in Makefile.sm.
• Added an explanation about the parameter routines
• Added an explanation for this currently empty header file
• added an extra link test to check if MPI_CC + MPI_LIBS can resolve pthread symbols to avoid repeated -lpthread during linking
• Added another debug message
• Added another error code shorthand macro
• Added blank for readability
• Added blank line for better readability
• Added cart creaet error detection test
• Added checks for pthread_cleanup functions, these may or may not be useful, depending on whether these must be lexically scoped
• Added coarse-grain thread lock support in MPI logging.
• added code to remove .exe from cygiwn's program name in forming clog2 and slog2 filenames.
• added code to the Fortran MPI_INIT and MPI_INIT_THREAD routines to extract command line arguments and pass them to their C counterparts. the globus device needs this since some vendor MPI implementations require the command line arguments (like MPICH-GM)
• added commands to generate MPE's manpages and htmlpages
• added comment
• Added comment about unused include base makefile feature on simplemake
• Added comment for eventual hook for using libatomic
• added comment re: removal of MPIU_THREAD_SINGLE_CS_ENTER/EXIT from MPI_Init
• added conditional macro HAVE_MPI_INIT_THREAD
• Added datatype and MPI_Gather changes to the CHANGES file.
• Added debug macros for the send queue
• Added definitions for *_FILENO -- defined in <unistd.h> in unix -- in mpe_logging_conf.h
• Added definitions for NMPI_Type_create_struct.
• Added extern of the MPID_Grequest_class_direct/mem for the no weak symbols case
• Added extra instruction on configure/make/install with different shells.
• Added fix for truncating short messages
• Added fixme on the barrier in finalize
• Added fixmes
• Added function prototype
• Added get_range to param
• Added global variables for MPE logging instrumentation
• Added help messages for MPI_CFLAGS and family.
• Added hook for loadable extensions in simplemake
• Added include <string.h>
• Added info about g95 and stdin
• Added info on dll channel
• added info on optimization level
• Added memarena and dll tests to configure
• added memset() to initialize CLOG_CommIDs_t array after malloc and realloc to keep valgrind happy on 64bit machine.
• Added missing MPIDI_QUOTE definitions
• Added missing hyphen in prefix
• Added missing ierr arg to MPI_Abort call in test
• Added missing memory barrier tests for Solaris
• Added missing prototypes
• Added missing segment free
• Added missing set of remote size before loop
• Added missing unistd.h and test for HAVE_UNISTD_H
• Added missing void in function prototypes
• Added more caveats about dllchan
• Added more clean steps for the shared libraries with version numbering
• Added more comments about the reason for not using -fno-common
• Added more connection debugging support
• Added more debugging messages and fix formatting errors
• Added more diagnostics when check for stdio linking with Fortran fails
• Added more information about reason for failure
• Added more message packet debug code
• Added more verbose output and reduced the noise tolerance threshold such that summation of user and MPI noise is lesser than the error threshold.
• added mpe_callstack_ldflags support in mpe<compiler> -mpilog and -log
• Added mpi_base module name for compilers that make this upper case
• added mutex debugging for posix mutexes
• added new install header mpe_log_thread.h
• added new mpe_log_thread.h
• added noncontig send function pointers to vc. Noncontig send functions can now be overridden by channels on a per-vc basis by updating these pointers.
• Added note about stdin limitations in MPD
• Added note about using killall
• Added parm ameter description extractor
• Added partial support for extended F90, including basic module
• Added path for the header files included by mpiimplthread.h and mpiimpl.h
• Added path to mpe_callstack.h -- MPE Callstack implementation is NOT complete on windows
• Added pending_pkt to request
• Added pkgconfig support for ch3 device, as an alpha test
• Added probe-intercomm
• Added probe-intercomm source into Makefile template
• added pthread version of concurrent allreduce and send&recv (in each thread)
• Added query about dependentLibs
• Added quotes to the smpd module file name before passing to CreateService -- Refer Req3145 & MSDN doc of CreateService() for details
• added rendezvous function pointers to vc. Rendezvous functions can now be overridden by channels on a per-vc basis by updating these pointers.
• added rendezvous threshold value to vc structure. Channels can now override this value.
• added return 0
• Added runhalo and options for MPI RMA optimizations in the halo exchange tests
• added runtime thread check macros in a few places
• Added section on Performance Analysis using MPE
• added separater between subpackages.
• Added some C++ pt2pt tests
• Added some dll tests for Unix-style dlopen
• added some more MPE documentation.
• added static
• added struct name to generic packet typedef
• Added support for ABI versions in shared libraries
• Added support for EXTRA_LIBS
• Added support for ICOMM_WORLD, an internal comm_world
• Added support for POSIX environments for singjoin
• Added support for a private copy of comm work
• Added support for invoking memory arena validation on function enter and exit
• Added support for optional DLL libraries and eliminate duplicates when the same source file appears in multiple library targets
• Added support for the type_create_f90 functions
• added support to display statistics stored in CatgeorySummary in Legend windows, i.e. count of the drawables, inclusive and exclusive ratios of specific category type. Also added the support to automatically adjust the column width based the content of each column. The goal is to display every string and number shown in legend window without user to adjust the column width.
• Added support to nemesis for receiving multiple messages in a single cell (or buffer). Netmods can call MPID_nem_handle_pkt to receive one or more messages from the same buffer. Bugs in the CH3 packet handler code were also fixed.
• added support to produce better error message, i.e make use of MPE_CallStack_xxx to produce backtrace when error is detected.
• added support to set enable_safePMPI=no when enable_threadlogging=yes to satisfy the limitation of coarse-grained thread support that some global variables should be made thread-specific.
• Added targets to clean any local summary.xml files
• Added temp stub for pmi2, this probably should be completely rewritten, but must be checked in before the move to svn
• Added test for ext_procs existence for network polling
• Added test for partial comm when building a cartesian communicator
• Added test of creating an intercomm from an intercomm with MPI_Comm_create
• Added test of creating many communicators with comm_split, using MPI_UNDEFINED for some of them (looking for a leak of context ids)
• Added testing target for comm tests
• Added tests for __func__ and export of thread level and type
• Added tests for specific atomic memory operations
• Added the beginnings (not ready yet) of a dynamically-loadable channel
• Added the performance tests into the basic tests
• Added the second test program from Mathworks
• Added unpack test and cleaned up code; improve reproducibility of measurements (more work is needed on shared clusters)
• Added upcall test for dlopen test
• Added use of standardized parameter routine to Finalize for handling the memory dump option
• Added warning about parallel make
• added/exported MPE_CLINKER and MPE_FLINKER for example Makefiles
• Adding include path for MPE callstack -- fmpe config
• Adding optimization support for the IBM xlc/xlf compilers.
• Adding project dependencies to the MPICH2 solution - Fixes the build order building projects according to their dependencies
• Adding support for soname
• Adding the generic path --using JAVA_HOME env variable-- for the java include files
• All the dataloop code in the ROMIO tree now builds. Will bring changes backto MPICH2 side next and verify correctness before beginning to use in ROMIO.
• All threads need to be joined in before calling MPI_Finalize.
• allow thread_serialized
• allow toolbar's Yaxis expand button to fully expand the Yaxis labels one level down and be immediately fully viewable in the viewport
• Allows only the rank 0 process to print out "Start ... Checking" and "Done ... Checking" message.
• Also alloc_mem in Fortran
• Also fixed Fortran versions of Accumulate, Get, and Put
• Another fix for using interface instead of external for the error handler functions in the tests
• Apply patch for totalview from: peter.thompson@etnus.com
• assigned meaningful colors to almost all MPI/IO calls.
• Attempt to use killall to remove any stdintest or stdintest2 programs that are orhaned by bugs in the process manager
• Avoid integer overflow in test on large systems
• be careful when renaming for shared libraries: match '.a' only at end of line
• Beginnings of supprot for dynamically loaded libraries
• Being removing old and unused rdma support from shm channel
• Bug fix for non-contiguous receive of large messages.
• Bug fix of sorts -- for some reason MPI_COMPLEX is a derived type but hasno heterogeneous dataloop. Until I know where it is created, I need to notabort when I hit this case.
• bugfix (untested) for bug report #3055 that when WRAPPERS is built with MPICH2 but fortran is disabled (e.g. no valid fortran compiler found) by MPICH2 configure so SIZEOF_MPI_STATUS is undefined. The correct thing for WRAPPERS to do is to disable building libmpe_f2cmpi.a and the related fortran link/run tests.
• bugfix for MPI2's use of MPI_Allreduce
• Bugfix for shm and ssm channels. Added missing read and write memory barriers for x86, and missing volatile in packet structure
• bugfix in atomic inline asm
• bugfixes in code for receiving data in packet handler
• By default, VS includes a new proj to all configurations -- removing x64 installer from all other configs
• C++ style comments don't work with --enable-strict
• Cache IP addresses when checking for unique hosts if starting onlyone mpd per host (the default).
• can't free GROUP_EMPTY
• cast fixes for files greater than 2GB
• cast the kvsname for the error message output
• Catch case of sending to pmiSock already closed.
• Catch stderr in ps checks
• ch3:sctp 1.0.6 updates
• ch3shmRelease requires sock.lib created from ch3sockRelease config, and not ch3sockPDebug
• Change dbg level to typical and add two Panic printfs (remove before release) in places where the code fails to queue an incomplete message
• Change debug class for one macro and add a MPIU_DBG_MSGPKT macro
• Change ping to handshake in test.
• Change the error message utility to use a shorter line length for each individual message in the ring, allowing the total message (of size MPI_MAX_ERROR_STRING) to be longer than any individual message. This seems to be a good comprimise between allocating enough space for the message ring with printing detailed individual messages
• Change the test for arrays in the shell to handle the case where the shell execute neither branch of the test
• Changed CVS host from harley to shakey
• Changed Executable output path from .\Release\bin to mpich2\bin
• changed coll buf size to 16 MB
• Changed locn of mpe_log_thread.h[.in] from src/wrappers to src/logging
• changed mpich2prereq to use channel name, rather than device_args, so things like ch3:nemesis:gm work
• changed packet handler interface. Packet handlers now take the receive buffer size so they can directly receive data from the buffer if its already available rather than passing back an IOV and have the channel do it
• Changed recv_dict_msg to loop around receiving 8-byte len.
• Changed sendflood timeout to 900 seconds.
• changed size of generic packet to be defined as x number of ints and y number of pointers to better reflect how the packet size changes on different architectures
• Changed the configuration, from Release to ch3sockRelease, for mpich2s project in the solution - All projects which depend on mpich2s require mpich2socks.lib which is created by ch3sockRelease configuration (and not Release configuration)
• changed the default --enable-threadlogging=yes.
• changed the default threadlogging to NO
• 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.
• Changed the mpi_errno handling to the new format.
• Changed the poll function to not return true for SEND when there is no data to send.
• changed top-level configure to make MPI_THREAD_FUNNELED the max thread level for channels which do not specify one
• changed version number, updates to readme and release notes
• Changing the newtcp network module to use the new communication format, i.e., the network module maintains its own buffers and does not rely on cells.
• Check existence of key before using.
• Check for ECONNRESET in send_dict_msg
• check for errors
• Check for src/util directory before assuming that it is present
• Check if the errcodes are not null before broadcasting them. If the application specifies MPI_ERRCODES_IGNORE, it'll be NULL.
• check return staus of MPI calls
• Checked for the non-portable POSIX_MUTEX_ERRORCHECK_NP before using it, we need to fix these thread files in general (it makes no sense to define macros for these low-usage, non-performance-critical operations such as create mutex)
• Chg if to elif in handling exceptions during send_dict_msg
• Chg pass to break in handling PIPE errs in send_dict_msg
• Chg word 'ping' to handshake in error msg to avoid confusing users.
• Clarify reason for strange loop
• Clean up error jump
• Clean up fixme on max procs in intercomm create check for duplicate processes
• Clean up includes for Intel compiler warning
• Clean up intercomm create context id
• Clean up replaced file string
• clean up small amount of leaked memory. reported by Lisandro Dalcin(MPICH2 req #3570)
• Cleaned out old unused defs
• Cleaned out unused defs
• cleaned up 64-bit warnings bu changing int into MPIDI_msg_sz_t
• cleaned up 64bit warnings in debugging statements
• cleaned up thread definitions for portability
• Combined polling for incoming connections and data into a single poll function.
• comment out debugging statement
• commented out numerous printf() output
• commented out performance tests for release
• Commit support for dynamically loaded channels for the sock, shm, and ssm channels
• compilers.
• Configure script now displays each unhandled definition or @foo@ only once. The script now also displays the number of unhandled definitions/@foo@s and the total number of instances
• consolidated all MPI_Abort()s to CLOG_Util_abort which will use MPE_CallStack_xxx to produce a backtrace before MPI_Abort().
• Correct comment
• Correct error in --totalnum
• Correct problem of not capturing stderr when using subprocess module in MPDTest.
• Correct several small problems including gerpping for PATH value from printenv.
• Corrected the estimated receive size for intermediate nodes.
• corrected typo in color name.
• corrected typo pointed out by Rajeev.
• Correcting typo - replace the complete template string @GNUC_MINORVERSION@ with 0
• cosmetic change
• Create mpe_callstack_conf.h
• Created a new separate INSTALL.cross that explains the 2 step cross-compilationprocedure and how to use cross_spec subpackage. Updated the version number,INSTALL and README to sync with each other.
• Creating MPI CXX Interface lib for gcc
• Datatype content setting was being repeated twice causing a memory leak.
• declared global variable MPIDI_CH3I_progress_completion_cond as static
• Define *_FILENO on windows -- defined in <unistd.h> on unix
• Define a testandsetifzero to significantly clean up the busy lock code
• Define mpishared.h to contain just the basics of the MPICH2 implementation environment to simplify building stand-alone modules, such as dynamically-loadable channels. In support of that, move thread variables from MPIR_Process into a new MPIR_ThreadInfo global variable
• delete unused variable
• deleted MPI_ from MPI_FFLAGS
• detect and encode errors when MPI_Grequest_complete fails
• disable check for global symbols
• disabled build of romio/common for the 1.0.6 release
• disabled checking of pthread.h and libpthread.a(so) when --disable-threadlogging is used.
• Disabled showing include files for segment_ops.c
• do the right thing when too many operations pending
• Doc for "-mapall" option in mpiexec
• Documentation improvements, additional error detection, improved handling ofpairtypes when used outside of MPICH2, bug fix (would recreate dataloops inall cases, losing memory).
• don't complain if GROUP_EMPTY is freed
• Don't need to allocate the temporary buffer at the root when the root is rank 0.
• don't use malloc directly; rather, use the wrappers
• Due to potential blocking of MPIDI_PG_GetConnString() the state of VC is now *_CONNECTING in MPIDU_Sock_connect(). Queueing requests when VC is in *_CONNECTING state. Changing state of VC to *_CONNECTING before invoking MPIDU_Sock_connect() from within ssm
• Dynamically allocating a buffer in a structure. This allows malloc to try to keep it 8-byte aligned; avoids large performance problems in datatype communication.
• Eliminate duplicate IPs for a single host from gethostbyname_ex; keep unique ones.
• Embed manifest file ONLY IF present
• Enable -O2 optimization by default. This optimizes, gcc, intel, Portland, Sun and Absoft compilers. More optimization flags might need to be added for other compilers.
• Enable the use of short eager packets (payload in the header). This passes all MPICH2 tests except disconnect_reconnect3
• Enhance test/mpid test for dynamic libraries
• Enhanced manyconnect to make sure that the test program exists
• Enhanced the IO async test
• Ensure ar, ranlib, and ENABLE_SHLIB are set
• Ensure mtest.o is in the allocmem link
• error checking for generalized requests
• exported all variables in cross_spec.txt so all subpackages can see them
• extracted out contents access so that code may be made more portable later.
• file detection if we are contig in memory.
• File renaming prior to extraction for use outside MPICH2. Sorry!
• final checkin of dataloop changes from the ROMIO side.
• Finish fixing intercomm create from intercomm in Comm_create
• First cut implementation of the Infinipath PSM module within nemesis
• fix James S Perrin bug where error by one process in ADIO_FileSysType_fncallcauses hang when everyone else does an allreduce.
• fix a couple small grammar issues
• Fix cart create to correctly invoke the error handler; on error, make sure that both cart_create and graph_create return null communicators
• fix compile errors if AIO has been disabled or detected not to work
• fix configure test to do the right thing if we can't find aio.h or sys/aio.h.Generaly only a problem if --enable-strict passed in.
• Fix debug class and check error return from post_close
• Fix error in setting up the ring in the shared pointer tests
• Fix for MPICH2 req 3107.
• Fix for alignment errors on 32-bit systems. We need a better fix for this.
• fix for disable-sharedlibs
• Fix for length of dirname in Strncpy
• Fix for moved thread items
• Fix for naming the shared memory segments plus making that routine more robust (never pass in an array without its length)
• fix for overwriting homogeneous dataloop with heterogeneous dataloop in pairtype code.
• fix for pg_destroy after finalize
• Fix for struct ifconf with POSIX selected on OSX
• Fix for xargs that do not support the -r option
• Fix formatting of routine to match man page generator requirements
• fix goofed up for loop
• Fix minor bug related to ifhns when using -machinefile option.
• Fix minor typo
• Fix miscellaneous memory leaks and formatting problems
• Fix misspelling in arg-with help text
• Fix problem with MPICH_INTERFACE_HOSTNAME being passed incorrectlyto spawned managers.
• Fix shm device for new generic packet
• Fix some formatting
• Fix storage leak in thread-multiple case for threads other than the main thread
• Fix test to allow err_pending
• fix the call to poke
• fix the statvfs f_basetype check for solaris
• Fix to Comm_copy for size < old size
• Fix to probe routines for the new intercomm context id code
• fix typo
• Fix typo in error check
• Fix typo in state names
• Fix typo in test for done with finding a buffered send segement
• Fix typo in test makefile
• Fix use of address-sized ints in RMA and use only local include files to avoid problems with directory separators in Windows
• fixed LaTeX bugs
• Fixed a bug in the calculation of the buffer size for gather.
• Fixed a bug in the initialization of IOV length.
• Fixed a compilation error.
• Fixed a few compilation warnings.
• fixed a few warnings
• Fixed a typo in setting the socket state.
• fixed aint
• Fixed bad tests for correctness and improved error reporting
• fixed blatant error pointed out by [MPICH2 Req #3504]
• Fixed broken formatting
• fixed bug in eager noncontig function. Size of pointer to header, rather than size of header was being passed as the header size
• fixed bug in error format string code when an invalid datatype is passed
• fixed bug that appended ":$(TRACE_SRCDIR)" to VPATH without checking if VPATH has been set"
• fixed bug where MPI_Recv was posted for size larger than receive buffer
• Fixed comm_split to handler intercommunicators
• fixed comment
• fixed comment.
• Fixed compilation warnings with ssm.
• fixed compiler warning on 64bit machine
• Fixed displs var in test
• fixed error class
• Fixed formatting in file
• fixed fortran configure test when no fortran compiler is found in user's PATH, i.e. can't preset F77 with MPICH2+MPE is built
• Fixed header check for elan
• Fixed headers
• fixed htmldir and mandir to be consistent with mpich2's
• fixed improperly declared variables in test programs. A variable called aint would be declared in an .h file, but the test program assumed the variable would be called asize, and v.v. Implicit none was added to all programs/subroutines/functions to catch any such errors in the future.
• fixed incorrect $< in explicit rule (should be $?)
• fixed incorrect reference of liblmpe.a (should be libmpe_f2cmpi.a) in comment
• fixed incorrect type cast of signed int to unsigned int in Preamble (potentially a problem).
• fixed info and error_codes
• Fixed misnamed extension in Fortran90 test
• fixed missing nest_incr and _decr
• Fixed missing ref to kvsname (used to work by accident because of where the field was placed)
• fixed missing use of aint
• fixed prototype for 64-bit platform
• fixed recv buf size larger than allocated memory
• Fixed several memory leaks in the datatype code.
• fixed small typo
• Fixed socket buffer size handling so that it does something and is controllable by the user; also added some connection debugging support
• Fixed some bugs in receiving data directly as non-contiguous datatypes in gather.
• Fixed some compilation errors due to type-casting problems.
• Fixed syntax error -- extra parenthesis
• Fixed tests on undefs cpp variables
• Fixed the bug raised by Sam Lang. i.e. fixed the error message system to emit the queued error message after InfoBox(Drawable).toString() is invoked in the case that the user byte arrary is NULL but the user format characters are Non-NULL, both when slog2print and jumpshot are used, different exceptions occur but same cause.
• Fixed the format size used for displaying debug messages.
• Fixed the handling of the second arg in Win_create
• fixed the invalid CLOG_CommIDs_t* handed out by CLOG_CommSet_t's table[] due to use of realloc() in the CLOG_CommSet_t's implementation. Added new member IDs4world and IDs4self in CLOG_CommSet_t so to remove global variables CLOG_CommIDs4Self and CLOG_CommIDs4World. Updated commIDs in log_mpi_core.c when add_intracomm() and add_intercomm() (which may have invoked realloc in table[]) in log_mpi_core.c.
• Fixed type for DISPLACEMENT_CURRENT
• fixed type_extent type to MPI_Aint
• fixed typo
• fixed typo in -y command line options
• Fixed typo in stdio with Fortran and Fortran 90 test and added diagnostics to the Fortran 77 version
• fixed typo, bug, and improved content.
• fixed typo, req 3156
• fixed undeclared variable in fortran test
• Fixed use of full mpiimpl.h in rlogtime.c by including only the necessary headers
• Fixed warning of less-than within a verbatim block (where it is ok)
• fixed warnings
• fixed x86 high resolution timer issue re rdtsc serialization. Configure now tries rctscp, then cpuid;rdtsc and finally rdtsc.
• Fixes for DLL support and added doxygen config file
• Fixes for MPICH2_INCLUDE_FLAGS
• Fixes for compilers with CPP processors that require (as the standard does) that the # begin in the first column
• Fixes for shared memory for OSX (more are still needed)
• Fixes for the Fortran allocmem test
• Fixing compilation warnings.
• Fixing some compilation warnings.
• fixups for tiling of types with LB and UB (think |---XXXX----| ). Thanks,Heidelberg!
• fn_exit should be called when MPICH_IS_THREADED is defined.
• 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
• Formatting fixes
• free the correct pointer
• Fully initialize the vc in vc_init
• Further fixes to the eager short support
• Further improvements to dims_create; now divides 12 processes into 4 x 3 rather than 6 x 2
• further minor updates
• GenBuild(): Suppressed processing of Makefile variables
• get the #define protection correct for two-argument aio_write
• give up on NFS asynchronous I/O operations
• Got rid of an extra copy for large messages -- sending non-contiguous data using two send operations instead.
• Handle MPD_PORT_RANGE in mpd.conf file; convert to MPICH_PORT_RANGE.
• Handle missing PATH in environment.
• Handle the case of a Fortran compiler that has reals and double precision the same size (e.g., g59 with the -r8 switch)
• Have handle_active_streams break and return for sigchld.
• Have sigchld handler in mpdman invoke mpd_handel_signal.Have handle_active_streams break and return for sigchld.
• Have to free the created datatype.
• Heidelberg!
• i give up. make a copy of the MPI_Request because the MPICH2 C++ bindings(apear to) clobber extra_state for some mysterious reason with certain C++compilers.
• If the user specifies any CFLAGS, append our default optimizations to it, instead of replacing our optimizations.
• Implemented the scratchpad approach in the VC structure for sock, shm, and ssm. Tested sock and shm on Mac (ssm still fails on Mac, did before, does now)
• Improve Dims_create
• Improved error messages when programs do not exit as expected
• Include CFLAGS on the CC line incase they are used to change the compilation behavior (e.g., for 64 bit or a different architecture)
• Include IPC only when USE_SINGLE_MSG_QUEUE is defined.
• Included windows header files containing defn of HANDLE type
• increase test's cb_buffer_size to match current state
• Increased the datatype size to reduce noise.
• Increased the number of iterations to reduce noise.
• Increased timeout for sendflood to 8 minutes
• Increasing the channel private space to 256 integers to accomodate the large VCs for SSM.
• initialize *error_code
• initialize uninitialized variables to make valgrind memcheck happy
• initinfo has no PMPI versions and is ponly part of the libmpich, not libpmpich
• Install shared library version of the C++ library
• into segment_ops.c instead.
• John Bent finds a bug with collective buffering of non-interleaved accesses
• just dropped in a comment so i wouldn't forget an idea
• Keep sshm up-to-date with CH3 changes
• Keep the ps check from reporting non numeric pids (that means that it failed to properly parse the ps output)
• Keeping mpich2 project uptodate with changes to ROMIO code - adding new files/removing stale files related to ROMIO and datatype code
• Keeping project uptodate with changes to ROMIO code - adding new files/removing stale files related to ROMIO and datatype code
• Latex formatting update
• little improvement in test.
• MPE Thread support for Windows - (1) Creating mpe_log_thread.h (2) Adding basic thread specific defns to avoid compiler errors. NOTE: These modifications do not (yet) enable thread support in MPE.
• MPICH2 1.0.6 ch3:sctp documentation updates.
• MPICH2 x64 Windows Installer
• MPICH2_VERSION & MPICH2_NUMVERSION now supported on windows
• MPID_nem_cell_t and MPID_nem_abs_cell_t weren't synchronized causing some of the nightly tests to fail.
• MT: Removed posting multiple connects for a VC. MPIDI_PG_GetConnString() can block and release the lock allowing another thread to try posting a connect for a VC in the VC_STATE_CONNECTING/VC_STATE_CONNECTED state.
• made threadID a local variable to avoid race
• Majority of dloop-reorg changes.
• Make ABI subversion at least 1 (osx rejects 0) and make a start on MPICH2_INCLUDE_FLAGS
• Make detection of a mismatch in the C and Fortran integers an error, since in that case we cannot build a correct Fortran interface
• make it a little easier to compile in the "dump all hints" code
• Make ssm use the common socket connection routine
• Make sure an address-sized int is used
• Make sure dependencies directories are rebuilt
• Make sure mpd_signum is reset before using in syscalls.
• make sure request_ptrs[i] isn't null before dereferencing
• Make sure that all MPI objects are freed so that any memory leaks will show up when MPICH2 is built with --enable-g=mem
• Make sure that special f90 type routines are included in the shared library version of libmpich.so
• Make sure that the lots-of-types test includes the mtest routines
• Make sure that there the results are not too noisy. Also, run the benchmark multiple times and take its average.
• Make sure that unexpected short requests are marked as having completed all communication
• Make the bash versions of the scripts more portable
• many aio feature tests failing incorrectly on darwin (os x) because its brokenaio.h requires sys/types.h
• merging generalized request extensions branch
• Minor changes to README re nemesis
• Minor cleanup
• minor correction in alignment tests.
• Minor fix for message
• Minor fix for reporting results
• Minor format cleanup
• Minor formatting fix
• minor language and style updates
• Misc cleanup
• mktemp() is deprecated -- using mkstemp()
• Modified CH3 functions to use MPIDI_msg_sz_t instead of int for consistency.
• Modified all CLOG2 code to use the newly added serial-MPI implementation when building libmpe_nompi.a. Using the serial-MPI implementation allows same clog2 files being generated when MPI is used or not, e.g. global communicator ID will be stored even when no real MPI is used. The added benefit is that this paves the way for clog2 files to be joined when no MPI is used.
• 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.
• Modifying encode_buffer and decode_buffer methods to make them non-static so that they can be used within other sections. (Used withing PSM module)
• More changes for moving to the scratchpad for a channels private part of a vc, this time for the ssm channel
• More clean up to allow dynamically loaded channels
• More cleanup, this moves packet definitions for socket open/close into the one file that makes use of them, and will allow the unification of the sock and ssm operations for setting up socket connections
• more cleanup.
• More compilation warnings.
• More fixes for Fortran name mangling detection
• More minor fixes for socket buffer len
• More steps torwards dllchan
• More steps towards building stand-along channels to allow dynamic loading
• More updates for intercomm split (not done)
• Move CH3_Business_card to where it belongs
• Move STATE_DECL to avoid problem with declarations in code
• Move declaration before non-decl
• Move misplacced mapping -> remote_mapping block and hoist constant-valued conditional out of loop
• Moved MPIDI_MSG_SZ_FMT definition to the top-level, instead of repeating it in each device.
• Moved coarse-thread support from src/wrapper to src/logging so MPE logging can make use of it. Replaced --enable-coarsethread by --enable-threadlogging + misc fixes.
• Moved inclusion of windows.h to use HANDLE type
• Moved the createf90types test to a different directory
• Moved the freeing of temporary buffer to be independent of different cases.
• Moved the get-business-card routine into ch3_init.c
• Moved the version information into a separate file to simplify the build of mpich2version
• moved win_free before buf free
• movement of the globus device code from the 1.0.4 branch to the trunk. patches to the core mpich2 framework are still required for the device code to function.
• Moving Windows fileno defns to mpe_logging_conf.h
• moving around contents support code.
• mpich2 project in ch3ssmPRelease configuration now depends on sockp.lib created by ch3sockPRelease
• mpidpre.h should be included before MPI_Aint has a format allocated to it.
• NFS can't use generic AIO: it has to lock around each operation to flush caches
• NFS doesn't need it's own version of nonblocking io
• nc_test uncovers bug in HDF5-specific workaround: don't override noncontig infile detection if we are contig in memory.
• nemesis newtcp bugfix
• nemesis newtcp implemented SendEagerNoncontig
• nemesis newtcp: noncontig bugfixes
• nemesis newtcp: the remainder of large partially received messages are now received directly into user buffer
• nemesis: added memory fences to shm lmt code
• nemesis: --enable-strict warning cleanup
• nemesis: added MPICH_INTERFACE_NAME support to new tcp module
• nemesis: added PG_Destroy VC_Destroy and InitCompleted functions to nemesis
• nemesis: added fixed double-free bug in lmt
• nemesis: added lmt interface
• nemesis: added missing errornames
• nemesis: added multithreading
• nemesis: added psm dir
• nemesis: added some documentation for the layout of the cell struct, and added a macro to (hopefully) clarify the padding added after the next field.
• nemesis: added support for MPICH_INTERFACE_HOSTNAME environment variable to tcp module
• nemesis: adding feature to allow netmods to send directly from user buf rather than from cell. Currently modifying GM.
• nemesis: bug fixes to eager noncontig, and added eager contig function pointers for netmods to override
• nemesis: bug fixes when compiling with thread-single and enable-fast
• nemesis: bugfix for enabled threads, but not runtime enabled
• nemesis: bugfix in progress loop
• nemesis: changed nemesis to use generic channel_private field in VC, rather than insert nemesis-specific fields in the VC
• nemesis: cleaned up some includes
• nemesis: fixed compile bug
• nemesis: fixed fcname macro
• nemesis: fixed hang bug in nemesis progress engine. Added code in blocking receive progress to check whether network module completed a request.
• nemesis: fixed huge races in intranode lmt code
• nemesis: fixed nemesis to use segment_ptr
• nemesis: fixed order of 63/32-bit library paths
• nemesis: fixed performance bug in newtcp module: If you want to enable NODELAY, you should set it to 1, not 0 :-)
• nemesis: fixed typos
• nemesis: gm bugfixes
• nemesis: gm module bugfixes
• nemesis: in requests, replaced ch.iov_offset with dev.iov_offset
• nemesis: intermediate bugfix for thread progress. Still broken when using networks.
• nemesis: lmt shared memory bug fix
• nemesis: minor cleanup
• nemesis: multithreading fix
• nemesis: newtcp module bugfix
• nemesis: newtcp module changes to send routines
• nemesis: newtcp typo fix
• nemesis: noncontiguous data is now packed directly into cells, rather than first creating iovs then copying from those
• nemesis: removed debugging code that was mistakenly checked in
• nemesis: removed mkstmp() prototype that may conflict with prototype in system header file
• nemesis: removed nemesis packet types from CH3 packet structure
• nemesis: removed netmod specific fields from VC. Using an anonymous buffer area instead that netmods cast to the appropriate private structure
• nemesis: shared memory is now only allocated when needed (i.e., when there is more than one process on a particular node)
• nemesis: singleton init no longer tries to access pmi
• nemesis: updated LMT interface to allow more flexibility
• nemesis:blocking fixed bug in init code where some nodes had only one proc and others had more than one. Those with one node werent calling the same number of PMI_Barriers
• new subdir to work with (dataloop).
• Not closing the thread handle after creating the thread because we need to wait on it later
• Now using fixed size scratch pad for channel specific data in Virtual Connection structure for ssm
• Now using the global_mutex in MPIR_ThreadInfo global variable instead of MPIR_Process var -- syncing with changes for supporting dynamically loadable channels
• OS X wants sys/types.h before aio.h
• Only get the vc when you need to
• Oops. Fixed a compilation error.
• Oops. Forgot to remove some code that I commented out for testing.
• Oops. Some functions assume that the srbuf structure and the internal buffer inside it are contiguous.
• oops...reversing commit that was supposed to be applied to the mpig branch
• Optimized threaded MPI logging by moving the access of thread specific storage out of the global mutex block.
• PVFS_sys_create has a new parameter (layout) now
• Part of the way toward supporting intercomm creation with MPI_Comm_create
• pass the proper type to PVFS_sys_dist_setparam
• Patched Elan module (Patch provided by Ashley Pittman @ Quadrics)
• poke progress engine if request not complete
• Ported alltoall to windows
• Prevent multiple threads from trying to get the connection string for the same VC. SMPD does not handle this case and the program hangs. Now the current thread bails out if another thread is trying to obtain the connection info for the same VC
• Print if unexpected pending count; remove before release
• procedure and how to use cross_spec subpackage. Updated the version number,INSTALL and README to sync with each other.
• Protecting header includes in a macro. Otherwise, they give errors on windows.
• Provide more details about the reason for failing when the F77 and F90 compilers are judged incompatible
• Provide prototypes for the param-from-env functions
• Put back CS_ENTER/EXIT macros around MPI_Init and MPI_Init_thread. These are needed to fix a --enable-threads=multiple hang when spawning
• quiet "may be used uninitialized" warning
• rearrange some code for testing.
• Recursive spawns fail since the search path of the exec for the spawned process is set to the incorrect working directory. Setting the working directory of an MPI process to current working directory if not already set.
• Reduced the number of iterations of sendflood. Nemesis takes too long to finish with overprovisioning.
• Refactored eager noncontig send functions. Factored out a function, SendEagerNoncontig, which takes a segment, and sends the message with an iov. This function can be overloaded to, e.g., pack directly into a buffer.
• Reference the sock/shm VC_GetStateString routines only if USE_DBG_LOGGING is defined
• Release configuration should use the non-profiled version of the sock library, sock.lib, and NOT sockp.lib
• Remove PSM module for the release.
• Remove dead code, add struct MPID_Request to quiet compiler
• Remove line feeds from the UUID generated
• Remove temp debug prints
• Remove the stdin tests from the command tests
• Removed $? which may cause problem on some systems.
• Removed a part of the additional copy performed at the root node, if the root node is not rank 0.
• Removed bogus threadinit
• Removed dead code
• removed debugging code
• Removed debugging set -x
• Removed duplicate comment line
• removed ebx register from asm blocks so as not to conflict when compiled as PIC
• Removed exporting CLOG_CommIDs4World
• removed global variables CLOG_CommIDs4Self and CLOG_CommIDs4World and replaced CLOG_CommIDs4World by CLOG_CommSet->IDs4world and CLOG_CommIDs4Self by CLOG_CommSet->IDs4self. The change is needed to fix a bug that didn't take into account of change of CLOG_CommIDs_t* handled out by CLOG_CommSet_t's table[]. The CLOG_CommIDs_t* could have been changed when CLOG_CommSet_t's table[] has bee realloc'ed.
• Removed ib & sshm channels from the windows makefiles
• removed implicit null from spawnargvf.f and spawnmultf.f tests because the old sparc compilers don't have a declaration for iargc()
• Removed incorrect use of MPIU_THREAD_SINGLE_CS_ENTER/EXIT from MPI_Init and MPI_Init_thread
• Removed mpiexec project from the solution
• Removed post build event to copy dlls to windows system directory
• Removed printf for branch that is untested by MPICH2 tests because MPICH1 and Intel tests do cover this branch (and it seems to work). We need an MPICH2 test for short, noncontiguous messages
• Removed printf that was causing runtests to freak out.
• Removed references to unavailable sections and features
• removed simple2 from Makefile for the release
• removed simple2pmi from release
• removed sub-micro timmer codes.
• 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.
• Removed the combined send and recv poll functionality. Bigger changes are needed before it is functional.
• removed the old MPE logging calls with hardwired eventIDs.
• Removed the old MPE logging calls with hardwired eventIDs. Replaced with MPE logging calls that are compatible to MPE2. The instrumentation can be turned on by -DADIOI_MPE_LOGGING.
• Removed unused code
• removed unused variable.
• Removed use of strncpy to work around a failure in one compiler that did not like using strncpy without declaring it
• Removing developer libs from x32 installable
• Removing typo error when creating smpd.exe in 64-bit makefile
• Renamed change_state macro to CHANGE_STATE.
• Reordered the elements of a structure to maintain alignment.
• Reorganized process locks to make them clearer and to remove buggy code using Sun mutex (for example, used mutex_lock for unlock). There are still major problems here, in particular, the definitions that control the include file are not defined in this configure but in other configures, such as ch3/ssm and shm channels, and thus the build of the mpidu_process_locks.c file is not consistent with the use in the channels
• Replaced $(MV) by $(CP) so that the repeated make won't cause unnecessary compilation/linking.
• replaced --with-cross by --with-cross-spec and some cosmetic fixes
• replaced AC_CHECK_SIZEOF(void*) by AC_CHCK_SIZEOF(void *) and replace CROSS_SIZEOF_VOIDP by CROSS_SIZEOF_VOID_P to make them compatibile with MPICH2's cross-compilation env variables.
• replaced the test that need CROSS_SIZEOF_VOID_GREATER by the standard test AC_CHECK_SIZEOF of void * that requires CROSS_SIZEOF_VOID_P which would be compatible with MPICH2' cross-compilation file
• Reset the spin count when we yield.
• restored CLOG_COMM_TABLE_INCRE to the original default value
• Results seem to be very noisy. Trying increasing the number of elements in the vector to reduce the percentage variation.
• reversed order of Post and Start
• revert prior cast patch
• revert the John Bent fix from March 2007, which broke a different accesspattern. Commit this version, which seems to do the right thing, and even if anew situation comes up where it does the wrong thing, will still get the rightbehavior.
• Reverting back to static methods for now. Will need to convert them to non-static methods eventually.
• Reverting changes to the MPI_AINT format fix. This needs to be fixed after the release.
• Reverting the alignment issue for nemesis. We need a better solution for this.
• SMPD now correctly locks the global thread lock before using the MPID util functions
• sendrecvl performance test conforms to test output requirements
• Setting socket state to only readable during the RANKSENT stage.
• Shifting around dataloop code to make it easier to use in other projects.
• silience a warning if AIO not enabled
• Since NEX SX-8's cross compiler refuses to link with .no file, added extra code to move .no to .o right before final linking for all serial executable creations.
• small AIO cleanup: use the FAKE aio routines if aio disabled or detected to not work
• some bugfixes for --enable-fast case
• Some fixes for tighter type scoping for nemesis (more to do)
• split the MPE section into subsections that describe MPI automatic and MPE user-defined logging.
• ssm requires a huge amount of space in the vc
• swapped 2 sections to make it consistent with README.windows
• Switch Nemesis to use the MPICH2_INCLUDE_FLAGS and dev.pending_pkt; also portability fixes for strict POSIX compilations
• Switch ssm to use dev.pending_pkt instead of ch.pkt
• Switch to newer format loader options (all in one rather than split)
• Switched ssm to use the new util/sock function as part of device cleanup
• sync with the source
• sync with the updated source.
• take a pass at quieting some --enable-strict=posix warnings
• take two on cast patch
• Taking care of DOS control characters before sending command file to unix host
• temporary attempt to debug thread-safety stuff
• Test for failure to find valid kinds for Fortran Offset or Address kind
• Test that the C++ compiler actually works
• The configure help message for selecting the process manager mentions forker; should be gforker.
• The dist target will fail (instead of issue a warning) if there are no man pages included. Use doc=no to build a tar file anyway (but do not do that for a release)
• The sock channel expects VC to be UNCONNECTED to successfully post a connect on the VC. Changing VC to *STATE_UNCONNECTED before posting a connect on a VC for the sock channel
• The stylesheet is assumed to be in the same directory as the XML test output file for nightly tests
• The test was doing Testall a second time even if the first testall is successful. This is incorrect according to the MPI standard, since the requests are set to MPI_REQUEST_NULL as soon as the first Testall is successful.
• The test wasn't freeing the datatypes, causing MPICH2 to report a memory leak.
• Things seem to be working fine, so remove the numerous #if 0 conditions.
• tightened up the scope of definitions to move towards supporting a set of channel DLLs (as well as improving the programming practice)
• to MPICH2 side next and verify correctness before beginning to use in ROMIO.
• Too much noise on atlantis. Increasing the number of trials to get a cleaner result. This will increase the execution time of the test though.
• transp-datatype.c prints debug information that causes nightly tests to report an error.
• Try fork multiple times before reporting failure.
• Type cast void* to char* before adding an offset.
• typo
• Typos corrected
• update Jumpshot-4 Users Guide: statistics on count, inclusion and exclusion ratios; Process-Thread and Communicator-Thread ViewMaps on Threaded MPI apps.
• update ch3:sctp to compile with current repo
• Update for additional refs to a Fortran include in a different directory
• Update to build old pmi docs
• Update version & Author list
• updated
• updated MPI-IO limitations
• Updated all config.* files to include NEC SX-8 definition.
• Updated all the example Makefiles with MPE compiler wrappers, MPE_CLINKER/MPE_FLINKER, that allow new LDFLAGS added by other package, e.g. callstack.
• Updated ch3sock library configurations in ch3shmRelease, ch3shmPRelease, ch3ssmRelease, ch3ssmPRelease - all now point to the library built by ch3sockRelease or ch3sockPRelease appropriately
• Updated changes file to mention the compiler optimizations support.
• updated comment
• updated comment in sample code.
• updated comment to better document the relationships among realloc()/add_intracomm()/add_intercomm()/commIDs.
• updated cpilog_pack.c with example to do source file tracing using __func__ and __LINE__ and added Darius's sub-micro timer support with USEC_TIMING.
• Updated data.slog2 and arrow.slog2 to the latest slog2 format, removed all errtime*.slog2 which can be produced by textlogTOslog2.
• updated date and added threaded logging status.
• updated error message
• updated html files with added content about expanded legend table for categorysummary statistics, and section on performance analysis of threaded MPI apps and how to jumpshot with different viewmaps.
• updated installation instruction for SUN MPI cluster tools 7.
• updated the color for MPI_File_Open().
• updated the colors for underlying IO calls.
• Updated the default compilation options to be not added into mpicc and family.
• updated the distributed version of userguide.pdf with new content of expanded Legend window on hidden CategorySummary statistics, added documentation on how to use jumpshot of threaded MPI logfiles, and performance analysis of threaded MPI implementations.
• updated thread support info
• updated to sync with added documentation on threaded and multi viewmaps stuffs.
• updated version info.
• Updated version of test solution from 8.00 to 9.00
• Updated windows makefile generator, GenBuild() - (1) Enable correct extraction of source files from Makefile.sm in sock/src and util/sock (2) Add new path for shmproc.c
• updated with @COPTS@, i.e. CFLAGS, in creating static executable.
• updated with latest PAC stuffs from mpich2's aclocal_f77.m4
• updated with latest PAC_F77 tests from mpich2.
• updated with modifcation of MPE_CLINKER/MPE_FLINKER between build and install.
• Updates for the dllchannel. These fix problems with building in mpid/common/locks (the configure there needed to use MPICH_INCLUDE_DIRS) and in the builddll scripts (which needed to properly set that value before invoking the configure in mpid/common/locks). The fix further emphasizes the need to refactor the ch3 code to properly modularize (the locks routines are built several times, with *different* include files. This should be ok because of how those files are used, but it is not robust).
• Updates to handle "new" messages to/from pmi_simple whichwere added by Bill Gropp.
• updates, including moving pack/unpack functionality out of mpid_segment.c andinto segment_ops.c instead.
• Updating error names in errnames.txt
• Upgraded ch3sshm project for VS 2005
• Upgraded ib project for VS 2005
• Upgraded impgen project for VS 2005
• Upgraded minalignrlog project to VS 2005
• Upgraded mpi project for VS 2005
• Upgraded pmi project for VS 2005
• Upgraded printirlog project to VS 2005
• Upgraded printrlog project to VS 2005
• Upgraded prof option to compilation scripts
• Upgraded project for Intel Fortran 9.1
• Upgraded project for VS 2005
• Upgraded solution for VS 2005
• Upgraded version of test projects from 7.10 to 8.00
• Upgraded wmpiconfig project for VS 2005
• Upgraded wmpiexec project for VS 2005
• Upgraded wmpiregister project for VS 2005
• Use %p for pointer in debugging printf
• Use MPICH2_INCLUDE_FLAGS instead of CPPFLAGS for MPICH2 include paths to avoid problems with the way autoconf invokes and processes configure arguments, in this case, the CPPFLAGS variable
• Use TerminateProcess() instead of ExitProcess() in MPI_Abort() -- In the case of an MPI_Abort() using ExitProcess(), the state of the DLL/process may be inconsistent leading to a hang when calling the entry point of the DLLs during detach - refer Req 3475
• Use USE_RDMA_READV/WRITEV to control compilation of code that uses ptrace to access memory in another process. This is non-portable (ptrace has different implementations in different versions of Unix) and may also require special support from the kernel (to avoid possible security issues)
• Use a pointer to a segment, not a full segment, within the request. This is only needed for non-contig messages, and if necessary, we could combine and optimize the Segment_alloc with Segment_init
• Use only the first 2 args in Fortran error handlers
• Use proper type for address ints
• use the MPICH2 error messages correctly
• Use the common PAC_C_MEMATOMIC to find memory atomic instructions instead of separate, incompatible versions in each of the configures that need atomic operations
• Used datatypes to transfer large messages in a non-contiguous manner.
• used in the NFS code.
• Vector copy code is duplicated here and in dataloops. Getting rid of this copy, as it is not even compiled in.
• we weren't setting status correctly with generalized requests. rectify.
• When only the header is received, reset the recv_active field to 0; we will get more data later.
• whoops. debugging slipped in on last commit.
• With SMPD, MPIDI_PG_GetConnString() blocks on the progress engine allowing another thread to post a connect for a connection in the CONNECTING state. Changed the state of connection to CONNECTING before blocking on MPIDI_PG_GetConnString() & queued data when we find connection in the CONNECTING state
• Workaround for a bug in some versions of autoconf that discard trailing blanks in statements - this broke code that change the IFS
• Workaround for bug in pgcc compiler when -fPIC is selected

Files for each revision