search

neuronsimulator / nrn

NEURON Simulator
http://nrn.readthedocs.io
Other
406 stars 118 forks source link

Neuron 9 and multisplit #1983

Closed SteMasoli closed 2 years ago

SteMasoli commented 2 years ago

Context

Overview of the issue

Hi. After a distro upgrade I have recompiled the latest NEURON code, but now the following code does not work in setting up multiple cores with multisplit.

h.load_file("parcom.hoc") p = h.ParallelComputeTool() p.change_nthread(8,1) p.multisplit(1)

There are no visible errors but models run only on a single core instead of many.

Expected result/behavior

The models work just fine but they run only one core. The expected behavior is to use multisplit on all available cores.

NEURON setup

Minimal working example - MWE

https://senselab.med.yale.edu/ModelDB/showmodel.cshtml?model=266806#tabs-2 Golgi cell model, morphology_1. Comment cao CONSTANT in cdp5 mod_file nrnivmodl mod_files nrngui -python protocols/01_SS.py

Logs

No error log

nrnhines commented 2 years ago

I agree there is a problem. Version 8.2.1 simulates on my apple M1 in 17.5s (starts with 8 threads) and current master simulates in 55s. Will look into it.

nrnhines commented 2 years ago

The problem began with 9.0.dev-7-gbd2c7ac2f 'Replace pthreads with std::thread and friends (#1859)' and 9.0.dev-6-ga10145903 does not exhibit the problem.

nrnhines commented 2 years ago

Apparently this is not specifically an issue with multisplit as performance results for nrntest/thread are (note cache1 means cvode.cache_efficient(1))

hines@michaels-macbook-pro-2 thread % nrniv perf1.hoc           
NEURON -- VERSION 8.2.1 HEAD (c15906924) 2022-08-09
...
nt   cache0  cache1
1    1.79    0.9
2    1.34    0.49
4    1.1     0.28
8    0.66    0.24

whereas

hines@michaels-macbook-pro-2 thread % nrniv perf1.hoc
NEURON -- VERSION 9.0.dev-72-gb226aabd2 master (b226aabd2) 2022-09-16
...
nt   cache0  cache1
1    1.88    0.9
2    1.44    0.9
4    1.35    0.89
8    1.35    0.9
nrnhines commented 2 years ago

Here are the details about what is needed to run the above tests.

git clone git@github.com:neuronsimulator/nrntest
cd nrntest/thread
nrniv perf1.hoc
nrniv perf2.hoc

nt and cache1 are the important columns. nt is the number of threads, and ideally cache1 should get smaller by a factor of 2 for each line. The 8.2.1 results for perf2.hoc on my apple M1 are

nt   cache0  cache1
1    0.36    0.32
2    0.19    0.18
4    0.11    0.11
8    0.11    0.0999999
alexsavulescu commented 2 years ago

Via #1993 I get: image

alexsavulescu commented 2 years ago

Also:

bin/nrniv perf1.hoc 
nt       cache0  cache1
1        10.24   2.61
2        6.02    1.31
4        3.56    0.7
8        2.51    0.56
bin/nrniv perf2.hoc 

nt       cache0  cache1
1        1.11    0.93
2        0.63    0.49
4        0.4     0.31
8        0.27    0.22
nrnhines commented 2 years ago

1993 now runs the Golgi cell model mentioned above on Mac M1 in 17.5 s.