Brendan Myers
2017-10-05 21:39:27 UTC
Hi All,
I have a general question about Open MPI used with a benchmark (IMB). I have
a RoCE NIC and this device requires 16 queue pairs to achieve full bandwidth
(~97GbE) in our performance testing app. Is anyone aware if that many queue
pairs being required will be a problem for running Open MPI + IMB with
maximum bandwidth used. I don't recall seeing any way to set the amount of
queue pairs Open MPI will use to complete a job, which makes me think this
question may show my lack of understanding of Open MPI. If more information
is required to answer this question please let me know what data is
relevant.
Thank you,
Brendan Myers
I have a general question about Open MPI used with a benchmark (IMB). I have
a RoCE NIC and this device requires 16 queue pairs to achieve full bandwidth
(~97GbE) in our performance testing app. Is anyone aware if that many queue
pairs being required will be a problem for running Open MPI + IMB with
maximum bandwidth used. I don't recall seeing any way to set the amount of
queue pairs Open MPI will use to complete a job, which makes me think this
question may show my lack of understanding of Open MPI. If more information
is required to answer this question please let me know what data is
relevant.
Thank you,
Brendan Myers