Hui Zhou
Hui Zhou
> I need help of a CH4 expert to get this PR working for CH4. For ParaStation MPI we had to make some modifications in our PSP device layer so...
Interest? Yes. It requires significant work and possibly a redesign for ch4 to work. But I think we can add some conditionals to allow merging these codes if Parastation MPI...
test:mpich/ch4/gpu/ofi ✔️ except two cuda memory allocation errors.
@abrooks98 Was the gpu/ofi tests clean?
Recording sample of the test failures:  I don't think they are related to this PR, so I will merge it and figure out the test failure separately.
Here is the diagnosis: `tmpi` relaunches hydra using `-pmi-port` option -- someone need explain to me how tmpi works -- so `hydra` will set a `PMI_PORT` string for launched process...
If we assume `localhost` is universal and always work, we can just use that. This is for MPI processes to connect to `hydra_pmi_proxy`, which should always work in localhost.
@s-sajid-ali The failure is because hydra listens on the hostname returned from `gethostname`, but on Mac, the hostname is just a name and not necessarily attached to a usable network....
@s-sajid-ali Could you try another interface, e.g. `en0`, depend on what your ifconfig shows?
The mpi process cannot connect to the proxy server (`hydra_pmi_proxy`) on either address -- ``` Unable to connect to 127.0.0.1 on 49909 Unable to connect to 10.0.0.37 on 51414 ```...