Open garlick opened 4 days ago
Our PMI-1 client assumes KVS values may contain any character except LF (including space and =
), thus the value=whatever
protocol element needs to be the end of the protocol line.
I tried to document the PMI-1 wire protocol here before implementing it, and at one time got a little feedback from Pavan on it. But I guess when you're documenting somebody else's work, you are on thin ice.
Any guidance that would help us tighten up our implementation or correct that spec would be helpful.
I suspect the addition is unintentional, and the result of the refactoring of the PMI1/PMI2 protocol implementations into the libpmi convenience library. We will take a look.
Problem:
mpiexec.hydra
v4.2.0 (and reportedly v4.1.1) can no longer launch Flux.It looks like a PMI-1 wire protocol change has crept in to the hydra implementation around v4.1, where values are returned with
found=TRUE
appended to them? Our client checks the protocol version and it is still being returned aspmi_version=1 pmi_subversion=1
.Was that intentional?
TL;DR flux-framework/flux-core#6072