in mpi programs i'm seeing a high cost in a routine that calls
I thought maybe the stat call was a culprit, so here's a small modification that doesn't call stat. Couple problems here prompting the 'WIP' tag
- right now I just read into a "big enough" buffer. 1k holds a lot of addresses, but it might not hold every address
- related: de-serializing the buffer assumes full records, one per ssg member. what happens if I read into my fixed buffer a fragment of a ssg address?
- the ssg state might need some more information -- maybe a header saying how big the statefile is or how many records one could expect to find?
- does each record need a checksum or a magic footer at least? Some way that the de-serializer can say "well these N records were just fine but this last one was garbled or incomplete