Closed alwinm closed 1 year ago
Do the additional cudaDeviceSync
calls impact performance?
Do the additional
cudaDeviceSync
calls impact performance?
It's a good question. I haven't looked carefully but it's likely sub-ms or within noise. The more impactful part of the CPU_TIME flag is the MPI collection and synchronization. To squeeze out all the performance juice I would recommend leaving the flag disabled.
Overhead of timers, measured by placing a Start and End adjacent with nothing in between.
0.0052 ms
It looks like the MHD restart test is failing with this message
C++ exception with description "filesystem error: cannot rename: No such file or directory [/ihome/sam/jenkins/eschneider/eschneider/cholla-automated-testing/PR-291/hydro/run_output.log] [/ihome/sam/jenkins/eschneider/eschneider/cholla-automated-testing/PR-291/hydro/bin/tHYDROtMHDReadGridHdf5_RestartSlowWaveExpectCorrectOutput/run_output.log]" thrown in the test body.
This indicates that it can't find the run_output.log
file. I'm not sure if this is just a one off error or if the filename was changed or something similar.
All the tests are passing now. Looks like it was a one off issue
(number - 1) / TPB + 1
. This has the benefit of never launching 0 blocks, which avoids cuda bug.