was trying it out on a mac too and the mps perf was not too bad compared to the GPU on a G5g.

It looks a bit too much at first but it fixes so much of the crap ppl usually complain about.
It provides a sane way of customizing things.

really wanna try hyper optimizing the runs for these!!
GPU-aware MPI wasn't a thing when I was first doing biophis.
AAAND also wanna explore the patterns to reuse them in some other ML stuff I've been fucking around with.

for the scihubers
1. sci-hub.se/https://doi....
2. sci-hub.se/10.1039/C3CP...

didn't even get the chance to tmux detach.

Couple more runs and im consistenyl getting los ~160 ns / day.

So TL;DR same spot price as M8g and C8g (10x on-demand) and 10x performance gain on a G5g.

and this for the price of $0.111 per hr (current spot price).
$0.42 on-demand.

still rather slow though!
i think the 100 ns / day baseline I got it from running on the commet cluste from UofC at san diego back in 2013 lol

right now doing runs on a c8g.2xlarge spot: mpirun -np 4 /usr/local/gromacs/bin/gmx_mpi mdrun -deffnm nvt

look! an extra 4 ns / day when using 4 mpi nodes instead of 5

the EC2 module itself does all the gramacs install - plain 'ol openmpi - but it also creates some config scripts to mount an EFS filesystem and configure SSH access for all the MPI nodes you wanna spin up

the comments on the EC2 module outline my process:
1. run terraform and make it go brr
1. ssh
1. rsync files to and from, while ignoring some artifacts
1. tmux shit
1. and a nugget for finding up-to-date AMIs (using a wrapper cmd that retrieves creds for AWS from 1pass)

made a recipe for this to make benchmarking relatively simple and quick: github.com/seafoodfry/m...

shout out to fastplotlib, matplotlib would choke on this many data points

think that's rather slow... vaguely remember clusters from a decade ago would get you ~100 ns / day.
But considering that you can get ~100 ns for ~$10, guess it ain't that bad a deal. (using spot instances.)