Kenneth E. Jansen
Kenneth E. Jansen
> > * There is a contiguous numbering of all nodes. I think I understand this for nodes. > * each element topology will have one contiguous block Keeping it...
@jedbrown @jrwrigh feel free to describe anything that @brtnfld and others developing CGNS might need to know that I missed in the above description
Thanks for the advice. I did not realize that CGNS did anything differently when reading a file with m processes that was written by n processes when n is not...
Thanks again for the response. Our file is "flat" in the sense that it is a single zone and we are expecting all ranks to read a range of the...
Request for you to be added sent but no response yet so it might be a while. In the interim @jedbrown suggested `lfs setstripe -c 16 .` to set the...
In the spirit of push it until it breaks mode, @jedbrown suggested `-1` and this produced a hang with 192 nodes (each with 12 processes) reading a file written originally...
``` kjansen@aurora-uan-0009:~> grep "#0 " out192ReadHang..* |grep MPIR_Allreduce |wc 1496 10472 251977 ``` so that leaves 808 processes (12*192-1496) doing something else.
For reasons we are still sorting out, we seem to get 12 control processes as well but filtering these with what they seem to all be doing on BT #0...
Slicing the other way, here is where the first 200ish of the 809 processes that are not at the `MPIR_Allreduce` in case that tells anyone anything (I can provide more...
getting it down to a page of what I perceive to be the most likely suspects (and you can see who I have taken my eye off with the grep...