Consider using Forks for parallelism?

[Shians]

By default parallel::makeCluster() uses type = "PSOCK" for its parallelism. But changing the type to "FORK" is more efficient on non-Windows systems. I suggest wherever makeCluster() is used the code be changed to

if (.Platform$OS.type == "windows") {
    cl <- parallel::makeCluster(n_cores, type = "PSOCK", outfile = "")
} else {
    cl <- parallel::makeCluster(n_cores, type = "FORK", outfile = "")
}

This also significantly reduces memory usage on systems where forking is available.

With my quick testing of 300 cells with 5000 genes and ks = 1:10 I saw the PSOCK version taking ~90 seconds and the FORK version taking ~60 seconds.

The downside would be that forks do not parallelise across clusters.

[wikiselev]

Hi, many thanks for your suggestions and very sorry for a late reply. Could you please create a pull request with the suggested changes and we will merge it to master?
Thanks again,
Vlad

[gorgitko]

@Shians See #104

[Shians]

Very nice! Glad to see such a clean revamp of the parallelism.