[Hackathon] Add a temporary option for benchmark data.

[olupton]

Description
Adds an option CORENRN_EXTERNAL_BENCHMARK_DATA that we can point to a shared directory on Permutter/Ascent/BB5. A draft directory on BB5 is /gpfs/bbp.cscs.ch/project/proj16/olupton/nersc-gpu-hackathon-dec-2021, which exists (possibly out of date) at $CFS/ntrain9/neuron/nersc-gpu-hackathon-dec-2021 on Perlmutter and /ccsopen/proj/gen170/neuron/nersc-gpu-hackathon-dec-2021 on Ascent.

If this is available, building CoreNEURON will also create a special-core binary that can execute the channel-benchmark input data provided in CORENRN_EXTERNAL_BENCHMARK_DATA.

[bbpbuildbot]

Logfiles from GitLab pipeline #26230 (:white_check_mark:) have been uploaded here!

Status and direct links:

• ✅ spack_setup
• ✅ build:nmodl:gpu
• ✅ build:coreneuron+nmodl:gpu
• ✅ test:coreneuron+nmodl:gpu
• ✅ build:neuron+nmodl:gpu
• ✅ test:neuron+nmodl:gpu

[iomaganaris]

LGTM
Just one question. How long does this take to run? Should we also include ringtest in case it runs faster? Otherwise for validation of standard functionality I don't think it adds any benefit and we can keep only channel-benchmark

[kotsaloscv]

LGTM

[olupton]

LGTM Just one question. How long does this take to run? Should we also include ringtest in case it runs faster? Otherwise for validation of standard functionality I don't think it adds any benefit and we can keep only channel-benchmark

On CPU it takes ~8s (setup ~ simulation), on GPU it's ~25s (setup >> simulation, we should fix that 😅) on a Perlmutter node.

I didn't remove anything related to the ringtest, so it's easy to run that too if we want to.

[iomaganaris]

LGTM Just one question. How long does this take to run? Should we also include ringtest in case it runs faster? Otherwise for validation of standard functionality I don't think it adds any benefit and we can keep only channel-benchmark

On CPU it takes ~8s (setup ~ simulation), on GPU it's ~25s (setup >> simulation, we should fix that sweat_smile) on a Perlmutter node.

I didn't remove anything related to the ringtest, so it's easy to run that too if we want to.

By setup you mean the NEURON setup or loading the coredat files?

[olupton]

LGTM Just one question. How long does this take to run? Should we also include ringtest in case it runs faster? Otherwise for validation of standard functionality I don't think it adds any benefit and we can keep only channel-benchmark

On CPU it takes ~8s (setup ~ simulation), on GPU it's ~25s (setup >> simulation, we should fix that sweat_smile) on a Perlmutter node.
I didn't remove anything related to the ringtest, so it's easy to run that too if we want to.

By setup you mean the NEURON setup or loading the coredat files?

Loading the coredat files, which are about 1GB. I think the GPU one is slower because there are currently a lot of tiny unified memory allocations for Random123.

[bbpbuildbot]

Logfiles from GitLab pipeline #26660 (:white_check_mark:) have been uploaded here!

Status and direct links:

• ✅ spack_setup
• ✅ build:nmodl:gpu
• ✅ build:coreneuron+nmodl:gpu
• ✅ test:coreneuron+nmodl:gpu
• ✅ build:neuron+nmodl:gpu
• ✅ test:neuron+nmodl:gpu