Specifications

OFED+ Host SW
True Scale Fabric OFED+ Host Software
RN 7.2.2.0.8 February 2014
22 Order Number: H31512002US
IFS/
MPI
Major
MVAPICH2 jobs run between PCIe HCAs and Third Party or Intel
HCAs, may not complete successfully. The test may abort with
an ibv_post_recv error.
Any MVAPICH2 job attempted on fabrics with PCIe HCAs and
Third Party or Intel HCAs, must zero the MV2_USE_SRQ
environment variable as show in the example of the NAS CG
benchmark:
cd directory_containing_benchmark
/usr/mpi/gcc/mvapich2-1.7/bin/mpirun_rsh -np 4
-hostfile mpi_hosts \
MV2_USE_RDMA_FAST_PATH=0 MV2_USE_SRQ=0
./cg.B.4
IB Third Party/
Other
NEW
Major
Some applications using Platform MPI 8.1 or 8.2 over PSM
exhibit poor scalability or worse performance at larger core
counts.
Use the Platform MPI mpirun command with the -intra=nic
option. This means that PSM's shared memory communications
will be used instead of Platform MPI's native shared memory
communications.
Intel IB HCA/
SHMEM
Minor
SHMEM collective calls using PE subsets can hang. This problem
does not affect collective calls using the entire PE set, which is
the more common case. A collective call specifies the entire PE
set when the parameters are set as follows:
PE_start = 0
logPE_stride = 0
PE_size = num_pes()
None
Intel IB HCA/
SHMEM
Minor
The SHMEM API reductions on complex types are not
implemented:
shmem_complexd_sum_to_all
shmem_complexf_sum_to_all
shmem_complexd_prod_to_all
shmem_complexf_prod_to_all
None
IFS/
IBAccess
Minor
When a port is down and does not have a LID assigned,
clear_p1stats or clear_p2stats will fail against the given
port
None
IFS/
IBAccess
Minor
When using vFabric, the OFED saquery command may use the
wrong P-Key and timeout waiting for responses.
Intel recommends using the iba_saquery tool, which is
included with IntelIB-Basic or IntelIB-IFS. iba_saquery will
work properly when vFabric is configured.
Table 11. Open Issues (Continued)
Product/
Component
Severity Description Workaround