"datatype/large_type_sendrec" test fails on bgq

mpichbot commented 7 years ago

Originally by blocksom on 2014-05-01 09:34:44 -0500

The datatype/large_type_sendrec test fails on bgq with the following error:

runjob -n 2 --ranks-per-node=1 --block R00-M0-N10-64 \
       --timeout 240 --envs BG_MAPCOMMONHEAP=1 :     \
       ./large_type_sendrec
errors = 4294967296

This fails with 1 and 8 ranks-per-node and with the following configuration options:

"global" lock

../../mpich/configure --host=powerpc64-bgq-linux     \
                      --with-device=pamid            \
                      --with-file-system=gpfs:BGQ

"per object" lock

../../mpich/configure --host=powerpc64-bgq-linux     \
                      --with-device=pamid            \
                      --with-file-system=gpfs:BGQ    \
                      --enable-thread-cs=per-object  \
                      --with-atomic-primitives       \
                      --enable-handle-allocation=tls \
                      --enable-refcount=lock-free    \
                      --disable-predefined-refcount

mpichbot commented 7 years ago

Originally by jnysal on 2014-05-21 08:47:29 -0500

There are issues with PAMI sending data >= 4GB. I also found a few places in pamid where we are using "unsigned int" instead of size_t for the buffer size. So I'll change the milestone to 3.1.2 for now as we need a fix in PAMI too.

mpichbot commented 7 years ago

Originally by jnysal on 2014-05-22 08:31:09 -0500

Fixes for pamid device are ready. This allows the test to pass on BGQ. Hence changing the target back to 3.1.1. For power architecture, additional fixes are needed in the PAMI library.

mpichbot commented 7 years ago

Originally by Michael Blocksome blocksom@us.ibm.com on 2014-05-22 09:31:43 -0500

In 98b5e585a61a8eccbd0224b64c66c505fa5ddf0e: pamid: Allow message sizes greater than 4GB

Allow message sizes >= 4GB

Fixes #2076

Signed-off-by: Michael Blocksome blocksom@us.ibm.com

pmodels / mpich

"datatype/large_type_sendrec" test fails on bgq #2076