GitLab

[svn-r4061] 1) Adding mpir-mpioinit.c to windows makefile (2) Updating simple makefile for wintcp channel

We were not zeroing the attribute list for an object at
{comm,win,type}_free time.

Reviewed by balaji@.

[svn-r4057] Added support for environment variable RANLIBNEEDED which, if set to NO, will cause simplemake to omit the RANLIB steps

[svn-r4056] Enhanced the interface builders to handle multiple comments in a function prototype line

argument string becomes a part of the environment, so altering the string
changes the environment. In our case, we were freeing the environment. On
some platforms, this was not causing a problem (probably because free
was being lazy in deallocating), but showed up on octagon.

This should resolve ticket #453.

path. In that case, finding the path to the proxy executable is a little
tricky. We need a cleaner approach to finding where the proxy is; this
code is becoming quite hacky.

[svn-r4050] Fixed f77 and f90 build when the base f77/f90 compilers are naive compilers, i.e. not MPI wrapper compilers, by using the patch provided by David Gingold.  The patch resolves bug report req#425.  Also, fixed VPATH build for f77 and f90 support.

ignoring the return statuses of the processes and returning success to
the launcher every time. This commit should fix that part and some other
minor bugs in the process cleanup.

thanks for the help, Dave!

Reviewed by buntinas@.

No reviewer.

served to make the handle allocation tracing angry.

warnings. This happens when the launcher tries to kill the processes, but
the bootstrap server has already passed the signal to kill the processes
(e.g., ssh has such problems where the signal management is not reliable).

[svn-r4040] Handling unhandled defns in windows configure script. 1) No need for putenv decl (2) Disabling comm overrides, mad module rel defns (3) Enabling pipe collectives

[svn-r4039] 1) Moving wintcp_module to wintcp - to follow naming convention of nemesis channel dirs 2)Update windows configure scripts & proj files to reflect the renaming of files done in r4037 & this change. MPICH2 on windows should compile now

of the bootstrap server. Also changed the naming convention to match
the rest of MPICH2.

This change in itself does not do anything for runtime selection of
bootstrap devices, but we'll need this when we actually add that code
in the future.

[svn-r4037] This change breaks MPICH2 windows compiln - renaming files to follow convention in other channels - remove 'module' in the file names

[svn-r4036] Using (INTEGER *8) fortran type for offsets instead of (INTEGER (KIND=8)) to support g77 on windows - Refer ticket #276 for details

No reviewer.

[svn-r4031] fixes #402 (NPB IS hangs) and other allgatherv hangs.  Also narrowed down volatile qualifiers from entire struct down to individual fields so we wouldn't have to cast away volatile in some cases

'int' not 'int *' -- man that's one old (but very minor) bug.

This is partly a fix for tickets #222 and #449, although both
of those still need work from RobL.

Reviewed by robl@.

No reviewer.

These make testing for ref counting bugs easier.

No reviewer.

that directory in the executed command. Right now, either version should
work, but the newer model is better when we do hierarchical proxies. As
a side-effect, we do not have to use sh within ssh anymore, since we
have a single command executed now.

case of an abnormal exit, so that the runaway processes can be cleaned
up properly (based on the PID, instead of the executable name).

Algorithm: Each proxy keeps track of the PIDs of the processes it
launches and listens on a socket for incoming connections from the job
launcher. If the exit is clean, this socket is not used at all. However,
if the job launcher wants to kill the application (due to a timeout,
or an abort by another process in the application), a connection is
established on this socket and a message sent to the proxy to kill its
corresponding processes. We only support one command right now (KILLALL).

This should resolve ticket #447.