Pass on nthreads to xgboost functions in scDblFinder

[csoneson]

We recently had some cases where scDblFinder() would grab all the cores on the machine, and run very slowly. Passing the nthreads parameter to the xgboost functions (which I believe would otherwise resort to parallel::detectCores()) seems to improve the situation and it finishes quickly (although as far as I can see it still uses a bit more CPU than specified via BPPARAM). The code below was enough to trigger the issue (both with versions 1.24.10 and 1.25.4):

library(scDblFinder)
library(SingleCellExperiment)
sce <- mockDoubletSCE(ncells = c(2000, 3000, 5000), ngenes = 10000)
sce <- scDblFinder(sce, BPPARAM=BiocParallel::SerialParam())

Is this something that you have also run into?

cc @mbstadler

[plger]

Hi,
Thanks for raising the issue... no I hadn't seen this, can I ask whether it's on mac/win/linux, and what version of xgboost you're running?

[plger]

Looks like this also changed with xgboost>3, and might even depend on how the package was compiled... but as I see no downside to your edits will just merge, so you can forget my question...

[csoneson]

I saw this on Linux, with xgboost version 3.2.1.1

[plger]

if things look okay on the devel builds I'll port to release after

[csoneson]

Thank you!