Hello,
There is now an implementation using multicore-OCaml in the ‘domains’ branch.
People are very welcome to give it a try and share the speedup they observe,
especially compared to fork-based parallelism.
Thanks to @nilsbecker for having motivated me.
Regards,
F.
If you don’t use the domains branch, then parany is using fork-based parallelism.
If you want to use the domains branch, you need to install multicore-ocaml first:
opam switch create 4.12.0+domains
eval `opam config env`
Some benchmark using 8 threads to process 100,000 molecules (runtimes are in seconds):
| Chunk-size | Multicore-OCaml | (speedup) | Fork-based // | (speedup) |
|---|---|---|---|---|
| 1 | 63.344 | 1.19 | 64.383 | 1.17 |
| 2 | 57.195 | 1.31 | 60.649 | 1.24 |
| 5 | 53.905 | 1.39 | 55.235 | 1.36 |
| 10 | 52.568 | 1.43 | 55.172 | 1.36 |
| 20 | 52.888 | 1.42 | 55.237 | 1.36 |
| 50 | 52.961 | 1.42 | 54.260 | 1.38 |
| 100 | 53.940 | 1.39 | 52.917 | 1.42 |
| 200 | 52.486 | 1.43 | 51.946 | 1.45 |
| 500 | 52.597 | 1.43 | 51.026 | 1.47 |
One good thing is that the multicore-OCaml program is more portable (it doesn’t need the operating system to provide the fork() system call).
I’ll try to find another benchmark; this one is probably limited by how fast I can
read things from disk.