Designing a (much) smaller caburst for the CI

[iahepburn]

We have an experimental caburst branch based on the very simple pyramid mesh (forgive the confusing branch name). https://github.com/CNS-OIST/STEPS4ModelRelease/tree/rallpack2mesh
This mesh takes just ~ 2.5s per run on 8 cores. We currently have 2, 4, and 8 splits for the mesh for STEPS3 runs, which are good numbers for the CI. I suggest 8 cores is optimal: https://github.com/CNS-OIST/STEPS4ModelRelease/tree/rallpack2mesh/caBurst/mesh/split_8/steps3

I tested different scalings etc and I suggest using settings as currently in the model, i.e. a mesh scale of 1e-6 and making the 'smooth' section the first 1µm of the mesh, as demonstrated in here: https://github.com/CNS-OIST/STEPS4ModelRelease/blob/rallpack2mesh/caBurst/full/validations/pics/pyramid_STEPS3_V_100traces_scale1_smooth1.png

Now we need to do the full statistic on this model, as for the STEPS4 paper.

[cattabiani]

steps 4 with 8 tasks (and 1 node) takes 22 seconds per run. 100 runs = 36 mins. What are the specs of the CI? Can we run stuff in parallel? How many nodes do we have available? 8 splits => 8 ranks. How may cores per node do we have available?

[iahepburn]

why on earth does STEPS4 take so long? Have you pushed the code for STEPS4? I can take a look at it.

[cattabiani]

what do you mean "push the code"? The code is the same no? Back in the days it was the same. I just needed to run the script with a parameter for steps 4 or 3

[iahepburn]

does that work? I didn't think the mesh would work with STEPS 4, but apparently it does.

[cattabiani]

Ah yes it works. I didn't check the results in detail but it worked. I left steps split the mesh. I used the one in the root folder

[iahepburn]

OK- I will give it a try. You can go back to sleep!

[iahepburn]

STEPS4 is taking 3s per run on my machine with 8 tasks.

[iahepburn]

Just eyeballing it looks similar between STEPS 3 and 4, though obviously we have to do some proper statistics. Jules is going to test the runtime on our CI machine, just to make sure we don't have similar runtime worries this side.

[cattabiani]

sorry guys. As you can imagine I am a little busy on the closing of bbp. In particular we have a last-minute paper that we are trying to push through and that got priority as without bb5 the researchers will not be able to do more simulations. Just to say that I did not forget

[iahepburn]

We have this now, @cattabiani https://github.com/CNS-OIST/HBP_STEPS/pull/1265
Not sure if it's totally what you had in mind, but it is quick enough to be part of the CI, and hopefully sensitive enough to pick up any major issues.