Benchmarking
Last modified by paulapopa on 2022/10/04 08:19
TVB-INVERSION 1.0.0
Sampling priors - Remote execution
1. Execution times
In this section we provide some benchmarks regarding the execution times for the sampling priors step within tvb-inversion workflow, where we need to run a large number of simulations.
These have been computed on DAINT-CSCS HPC, on a single node, with different configurations for the number of simulations and the number of workers.
Model | Sim length (s) | Regions | Nr simulations | Nr workers | Execution time (hh:mm) |
|---|---|---|---|---|---|
| MontbrioPazoRoxin | 30 | 100 | 30 | 30 | 00:17 |
| MontbrioPazoRoxin | 30 | 100 | 200 | 20 | 01:08 |
| MontbrioPazoRoxin | 30 | 100 | 300 | 30 | 01:10 |
| MontbrioPazoRoxin | 30 | 100 | 400 | 40 | 01:18 |
| MontbrioPazoRoxin | 30 | 100 | 500 | 50 | 01:34 |
| MontbrioPazoRoxin | 30 | 100 | 500 | 55 | 01:30 |
| MontbrioPazoRoxin | 30 | 100 | 600 | 55 | 01:45 |
| MontbrioPazoRoxin | 30 | 100 | 600 | 60 | OOM |
| MontbrioPazoRoxin | 60 | 100 | 500 | 40 | 03:07 |
| MontbrioPazoRoxin | 60 | 100 | 500 | 55 | OOM |
2. Limitations
- Reaching the memory limit on the CSCS node.
- For a simulation of 30 seconds, we can fit 55 parallel workers in the available memory
- For a simulation of 60 seconds, we can fit 40 parallel workers in the available memory
- Reaching the maximum connections on CSCS could happen during a run
- connect once to launch the job
- connect multiple times during the monitoring step to check the status of the job
- connect once to stage out results