add compy (login and compute) and chrlogin to globus endpoints

[mahf708]

fix #274

• Especially important to review: a controversial workaround to guide the regex to compy if the the HOSTNAME is set to something starting with compy and the the fqdn is n*.local.

[forsyth2]

You make a good point, including the compute nodes for Compy. I think Perlmutter compute nodes also don't fit this pattern.

I usually run zstash from a login node but I suppose some may find it useful to run a large Globus transfer from a compute node.

[mahf708]

I was going to suggest at some point that instead of asking people to do screen we could suggest an alternative (that I personally prefer): submit the zstash job as a compute job. But first, I want to investigate what we can do make use of a full node (i.e., more threading, etc.)

[forsyth2]

@mahf708 The issue I discovered with using compute nodes is that sometimes the Globus transfers take longer than the max wall clock time (12 hours on Perlmutter, according to https://docs.nersc.gov/jobs/policy/). That is, I think we'd need to set up some sort of restart file system (as E3SM itself does) and use zstash --update.

[mahf708]

@forsyth2 I didn't test these changes "in production" yet. Would you like me to test them before we proceed or do you think your test suite will be able to able to handle them?

[forsyth2]

@mahf708 The only test that uses Globus is this one: https://github.com/E3SM-Project/zstash/blob/main/tests/test_globus.py#L182. I suppose we could run that test on both a Compy login node and a Compy compute node.

[forsyth2]

• Especially important to review: a controversial workaround to guide the regex to compy if the the HOSTNAME is set to something starting with compy and the the fqdn is n*.local.

I unfortunately don't know too much about the HOSTNAME issue. (I rarely run zstash from compute nodes).

That said, I ran python -m unittest tests/test_globus.py on login nodes and compute nodes for Chrysalis, Perlmutter, and Compy (using this branch) and it passes in all 6 cases. (For reference, I ran on compute nodes using salloc -n 1 on Chrysalis and Compy, and salloc --constraint=cpu -n 1 --qos=regular on Perlmutter.)

So, I think this is good to merge. Thanks @mahf708