Nikoli Dryden
Nikoli Dryden
@bvanessen Any updates on this? This isn't a blocking issue, but the error spam at large scale is annoying.
Well, that function is non-POSIX (hence the `np`), so we should probably just not use it. We can set affinity successfully via HWLOC.
One should be careful with too many threads though: processor cores are also a shared resource, and threads are not free.
We should, ideally, be flexible with regard to data layouts. For example, convolution in NHWC layout is preferred for using Tensor Cores. It may make sense to transpose layouts between...
This is actually a segfault that appears to be a bug in SpectrumMPI. The hang is because of our signal handler sucking. A simple reproducer results in a segfault on...
I did some work on implementing this when pysendfile is present to provide sendfile support. Some simple tests indicate that it works both when sendfile is present and when it...
This has been fixed in the latest SpectrumMPI release.
I was able to individually refresh every feed without a crash, hence my guess that it had something to do with initiating many refreshes at once. When refreshing groups of...