History of problems fixed
The following fixes apply to IBM PE updates for both Power Systems and System x, unless noted otherwise. All problem fixed apply to both AIX and Linux unless stated otherwise.
IBM PE 5.2.2.8 [May 2, 2013]
POE:
- Fixed a problem with buffer overflow with MP_LABELIO with a lot of messages.
- Fixed an issue with missing redirected STDOUT output.
MPI:
- Fixed a problem with a Fortran MPI IO file handle has the incorrect size.
- Fixed an issue with MPI_Sendrecv get 'wrong result' for small data types.
- Fixed a hang with MPI_Barrier() with large number of tasks on a job.
- Fixed a deadlock between MPI and LAPI.
LAPI:
HPCT:
PDB:
IBM PE 5.2.2.7 [March 28, 2012]
POE:
- Fixed a problem with tasks using sub jobs and task affinity.
MPI:
- Provide a way to tell the MPI library level being executed.
- Fixed an issue with programs trying to set IBM_largeblock_io.
LAPI:
- Fixed a failure during LAPI failover when the HAL port has been closed.
- Fixed a hang on IP/Ethernet and IP over InfinBand.
HPCT:
- Fixed in issue with incorrect event group counts.
- Provide support for hpm to gather Westmere EX hardware counters.
- Fixed an installation issue with HPC toolkit X Windows dependencies.
- Fixed an issue with HPCT when compiling Fortran code with optimization.
PDB:
IBM PE 5.2.2.6 [December 14, 2011]
POE:
- Fixed I/O buffer overflow and poe exit.
- Fixed a problem with large number of parallel threads with MP_TASK_AFFINITY = cpu:n or core:n.
- Fixed problems with MPICH2 style compiler scripts.
- Allow single node IP jobs to be checkpointed.
MPI:
- Fix a problem with MPI_Gatherv using too much memory when shared memory is used.
- Fix to allow calling MPI_Alloc_mem with size = 0 on AIX.
- Fix a problem with MPI_CART_SUB in Fortran giving a wrong answer when programs are compiled with '-O2' option.
- Provide support for MPI_File_f2c, MPI_File_c2f, MPI_Status_c2f and MPI_Status_f2c functions on Intel
- Fixed a problem with Fortran bindings for MPI_Init_thread().
- Fixed a problem with MPI_Bcast() with a user defined datatype.
LAPI:
- Fixed MP_MSG_API setting for Init sample.
- Fixed PNSD file descriptor cleanup with adapter affinity.
- Provide support for AIX 32-bit get_adapter_affinity.
- Fixed incorrect queue pair free on setup.
- Properly initialize HAL receive FIFO.
HPCT:
PDB:
- PDB support of idbc for Intel compiled binaries.
IBM PE 5.2.2.5 [Sept 2, 2011]
POE:
- Provide task affinity support on x86 platforms.
- Fixed termination issues when Cntl-C is done before job is fully started.
- Fixed a problem with the PMD log text containing binary data strings.
- Fixed a problem in free() when MP_EUILIBPATH was set.
- Fixed a problem in the cpuset_query command on the x86 platform, where
the wrong information for CPUs belonging to a core was displayed.
MPI:
- Fix a PE MPI library error so that applications that are compiled using
the mpiicc compiler script and MPICH2 mpi.h will get the correct value
of MPI_BSEND_OVERHEAD.
- Fixed a problem in the mpicc script compiling with the -lapi flag.
LAPI:
- Fixed a problem in the Mellanox device driver.
- Improved bandwidth on Mellanox.
- Provide support of bond0 device mapping on InfiniBand.
HPCT:
PDB:
IBM PE 5.2.2.4 [Aug 2, 2011]
MPI:
- Fixed an PE MPI internal error that causes MPI_PROC_NULL to be
converted to rank zero. This error causes an application that
calls MPI persistent communication functions with MPI_PROC_NULL
to hang.
- Recompiled the MPI MODULE (mpi.mod) file with IBM XLF V12 compiler
so that IBM PE can continue to support this version of Fortran
compiler. The MPI MODULE file is for used by Fortran 90 programs.
- Fixed an PE MPI internal error that causes the MPI_File_set_view
function to hang.
- Fixed several C string to Fortran string conversion problems.
- Fix a MPICH2 support problem on X86_64 nodes with Sles11 Linux.
An MPI program that is compiled with a MPICH2 compiler script using
MPICH2 header file can now run correctly using PE MPI library on a
X86_64 node with Sles11.
POE:
- The MPICH2 compiler scrips, mpicc, mpicxx, mpif77, mpif90 and mpiicpc
now support compilation of LAPI programs. These scripts will compile
a lapi program when they are invoked with the '-lapi' flag.
- Fixed a poe problem when it is given a LD_LIBRARPY_PATH value that
is longer than 1024 bytes. This problem causes poe to coredump
on a X86_64 node with Sles 11 Linux.
- PMD is modified to minize calling of LDAP functions. When LDAP
functions are called too frequently PMD may hang when handling
large job on an installation which stores user credential data
on an external file system.
LAPI:
- PNSD is modified to minize calling of LDAP functions. When LDAP
functions are called too frequently PNSD may hang when it is
runing on an installation which stores user credential data
on an external file system.
- PNSD is enhanced to handle heavily loaded system. Without this
enhancement large job start up on a heavily loaded system may
fail with this error message:
Error: 0031-250 broken pipe
& 0031-161 end of socket connection
- PNSD now supports 512 InfiniBand windows.
HPCT:
PDB:
- Fixed a PDB core dump when used on a X86_64 node with Sles 11
Linux installed.
IBM PE 5.2.2.3 [May 27, 2011]
MPI:
- Correct the MPI_VERSION and MPI_SUBVERSION definitions in mpi.h.
The correct MPI version and subversion supported is MPI 2.1.
- Fixing IBM_largeblock_io support, without this fix the MPI library
does not honor setting the IBM_largeblock_io MPI-IO hint to truth.
- Improve MPI_ALLTOALLV performance.
- Fix a MPI_Comm_disconnect function defect which causes this function
to hang when it is called from a multi-thread dynamic tasking job.
POE:
- Fix HPCCOUNT support without this fix running hpccount in user space
with any plugin other then the default and with 2 or more procs
fails with error: ERROR: 0031-686 Unable to broadcast device initialization
information to nodes
- Fix support for Openssh on Linux, without this fix a system will
have to enable RSH (.rhosts) for standalone poe to work.
LAPI:
- Add Mellanox Blue Flame feature to improve latency.
- Update IB support on AIX to support Mellanox QDR 4K MTU.
HPCT:
- Support has been added to access the hardware performance counters
on x86 class machines for the Intel microarchitecture codename
Westmere family of processors.
PDB:
IBM PE 5.2.2.2 [March 11, 2011]
MPI:
- The X86 Linux version of the 'mpfort' compile script has been fixed
so that it will find the libmpichf90.a archive file.
- The X86 Linux compile scripts -- mpicc, mpicxx, mpif77, mpiicc
and mpiicpc are fixed so that programs can be linked with the
libmpitrace archive file.
- On X86 Linux the MPI library will return the correct upper bound
value for MPI_TAG_UB (1681915906) when a program is using the
MPICH2 header file.
POE:
- Resolved an issue with additional task affinity support, when
the SMT mode is dynamically changed.
LAPI:
HPCT:
- Elminated a dependency on a binutils demangle routine.
- Add support to handle long function names that C++ tends to generate.
- Mpitrace now supports MPICH2 headers on X86-Linux.
PDB:
IBM PE 5.2.2.1 [Dec 10, 2010]
MPI:
- Added MPI C++ language bindings that are missing from the
PE 5.2.2.0 version. (This fix applies to IBM PE for Linux ONLY.)
- Fixed an internal error that will cause a 32bit program to hang
if this program is accessing a larger than 2GB file with MPI-IO
functions.
- When compiling MPI programs on an IBM X-Architecture system using
the gcc compiler, the IBM PE compiler scripts will include the
MPICH MPI header file and will link the exec to a MPI library
that is built by the gcc compiler.
- The 'mpfort' Fortran compiler script for Linux has been changed
to use 'gfortran' as the default compiler.
POE:
- Fixed a poe error that will cause a standalone job to hang when
the poe option "-task_affinity" is set to "cpu:1" or "core:1".
- Fixed an error that causes poe to hang when handling a job that
uses mix MPI and LAPI protocols.
- Poe has improved performance for 'co-scheduler' support.
LAPI:
- Fixed a PNSD hang problem on IBM X-Architecture system when the
system has enabled IPv6 protocol. This is a Linux only problem.
HPCT:
- Elminated a dependency on a binutils demangle routine.
- Allows handling of long function names that C++ tends to generate.
- Enable mpitrace to work on X with MPICH2 headers.
PDB:
Totalview Support:
IBM PE 5.2.1.6 [Sept 3, 2010]
MPI:
- An internal error causes an application which is making persistent
communication calls to be terminated with the "handle is active"
error message while the MPI_Startall function is attempting to
activate all requested handles. This error is fixed by ensuring
that during the execution of the MPI_Startall call, all requested
handles are not in active state.
- Fixed an error in the MPI_Irecv code that causes the return of
an invalid MPI request handle to the caller when MPI_PROC_NULL
is specified as the source of the receive.
- Fix the MPI_Type_dup function to make it conform to the MPI standard
POE:
- Fixed a coredump error when poe is trying to execute an application
with a path name with more than 256 characters.
- Fixed an error that causes the cpuset_query to return an erroneous
core count when the SMT architecture is SMT4.
- Update the poe.README file to add restriction for performing
Checkpoint/Restart on AIX Diskless cluster.
- Add the pm_delete_shmkey command to the Linux RPMs. It is called
by PMD, internally, to make sure there are no stale shm keys left
behind by existing tasks.
LAPI:
- Fix PNSD, on Linux, so that it will assign a correct IP address to
an IB adapter which has not been configured.
HPCT:
PDB:
Totalview Support:
IBM PE 5.2.1.5 [July 23, 2010]
MPI:
POE:
- Fixed two internal errors that cause the pmdv5 to coredump.
Linux only.
LAPI:
HPCT:
PDB:
Totalview Support:
IBM PE 5.2.1.4 [July 12, 2010]
MPI:
POE:
- SMT-4 mode support for Power 7.
- Various internal fixes.
Linux only.
LAPI:
HPCT:
PDB:
Totalview Support:
IBM PE 5.2.1.3 [June 15, 2010]
MPI:
Fixed a MPI_Cart_sub problem so that this function will comply with the MPI standard to
return a zero dimension Cartesian topology to all ranks during partitioning of a Cartesian
topology into subgroups if all of the remaining dimensions are false.
POE:
The poe command has made improvement to guarantee when a job has requested TASK AFFINITY on a Linux
system then each task is attached to a CPU set before execution is started. This fix is for
Linux only.
LAPI:
- Fixed an internal error so that LAPI can handle larger than 4GB messages.
- Fixed a problem that the nrt_status command is generating excess among of logs when it is
run on a node with running jobs.
HPCT:
PDB:
Totalview Support:
IBM PE 5.2.1.2 [May 13, 2010] - AIX only
MPI:
- Fix a mpxlf90_r, Fortran compiler wrapper script, problem that causes it get an erroneous _XL_VERSION when the IBM XLF
compiler is installed in a non-default location.
- Fix a mpi.mod, pre-compile header file for Fortran program, problem that can cause some Fortran program to fail compilation
due to erroneous function parameter check result.
POE:
LAPI:
HPCT:
PDB:
Totalview Support:
IBM PE 5.2.1.1 [May 13, 2010] - AIX only
MPI:
- Fixed data corruption problem caused by calling MPI_Wait in a single thread runtime environment.
- Fixed data corruption caused by MPI_Allreduce, when the application is run on a Power 7 node.
POE:
- Fixed memory leak problem in coscheduling kernel extension. The kernel extension has memory leak when using kernel level rset. With this, fix memory used is automatically freed at task termination.
- Fixed problem that poe does not check the validity of the -cmdfile flag. Also improved the error message when an input
error is detected.
LAPI:
HPCT:
PDB:
Totalview Support:
Installing a PE update for AIX