figGAM is not adhering to BPPARAM limits.

[dfermin]

Hello

I am trying to run the fitGAM command on a multi-core linux server.
The machine has 64 cores and I want fitGAM to only use 16 cores.

I've followed the vignette's suggested method for running in parallel but fitGAM() seems to ignore this and uses all available cores.

Here is my code

Dataset <- readRDS("cellStateData.rds")
Dataset <- as.SingleCellExperiment(Dataset, assay = "RNA")

Dataset <- slingshot(Dataset, reducedDim = 'UMAP', clusterLabels = colData(Dataset)$celltype, approx_points = FALSE)

workers <- BiocParallel::MulticoreParam(workers=16)
Dataset1 <- fitGAM(Dataset, verbose=TRUE, parallel=TRUE, BPPARAM=workers)

I'm running this on a 64 core RHEL server.
When it gets to the fitGAM line the command uses all 64 cores.
There are no previous commands that might use multiple cores.

I've tried different renditions of the BPPARAM argument all to no avail.
Any suggestions on how to fix this?

Thanks

[Alexis-Varin]

Hi, I think this is a problem from BiocParallel and not tradeSeq, I noticed the same behavior on any Linux machine I used, I would suggest using SnowParam() and see if it passes your parameters correctly.

[Crml23]

I'm running into the same issue. However, it doesn't seem to be an issue with BPPARAM. Even when parallel=F in the fitGAM call, it uses all available cores.

Here is the function call:

 sce_gam <- fitGAM(counts = counts, 
                    pseudotime = pseudotime,
                    cellWeights = cellWeights,
                    nknots = 9,
                    verbose = T,
                    parallel = F)

Any advice? Thanks.

R version 4.4.2 (2024-10-31)
Platform: x86_64-pc-linux-gnu
Running under: Ubuntu 24.04.2 LTS

Matrix products: default
BLAS:   /usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3 
LAPACK: /usr/lib/x86_64-linux-gnu/openblas-pthread/libopenblasp-r0.3.26.so;  LAPACK version 3.12.0

locale:
 [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C               LC_TIME=en_US.UTF-8        LC_COLLATE=en_US.UTF-8     LC_MONETARY=en_US.UTF-8   
 [6] LC_MESSAGES=en_US.UTF-8    LC_PAPER=en_US.UTF-8       LC_NAME=C                  LC_ADDRESS=C               LC_TELEPHONE=C            
[11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       

time zone: America/Los_Angeles
tzcode source: system (glibc)

attached base packages:
[1] stats4    stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
 [1] slingshot_2.12.0            princurve_2.1.6             tradeSeq_1.18.0             TrajectoryUtils_1.12.0      SingleCellExperiment_1.26.0
 [6] SummarizedExperiment_1.34.0 Biobase_2.64.0              GenomicRanges_1.56.1        GenomeInfoDb_1.40.1         IRanges_2.38.1             
[11] S4Vectors_0.42.1            BiocGenerics_0.50.0         MatrixGenerics_1.16.0       matrixStats_1.3.0          

loaded via a namespace (and not attached):
  [1] RColorBrewer_1.1-3      rstudioapi_0.16.0       jsonlite_1.8.8          magrittr_2.0.3          spatstat.utils_3.0-5    farver_2.1.2           
  [7] zlibbioc_1.50.0         vctrs_0.6.5             ROCR_1.0-11             spatstat.explore_3.3-1  S4Arrays_1.4.1          htmltools_0.5.8.1      
 [13] SparseArray_1.4.8       sctransform_0.4.1       parallelly_1.38.0       KernSmooth_2.23-24      htmlwidgets_1.6.4       ica_1.0-3              
 [19] plyr_1.8.9              plotly_4.10.4           zoo_1.8-12              igraph_2.0.3            mime_0.12               lifecycle_1.0.4        
 [25] pkgconfig_2.0.3         Matrix_1.7-1            R6_2.5.1                fastmap_1.2.0           GenomeInfoDbData_1.2.12 fitdistrplus_1.2-1     
 [31] future_1.34.0           shiny_1.9.1             digest_0.6.36           colorspace_2.1-1        patchwork_1.3.0         Seurat_5.1.0           
 [37] tensor_1.5              RSpectra_0.16-2         irlba_2.3.5.1           progressr_0.14.0        fansi_1.0.6             spatstat.sparse_3.1-0  
 [43] mgcv_1.9-1              httr_1.4.7              polyclip_1.10-7         abind_1.4-5             compiler_4.4.2          BiocParallel_1.38.0    
 [49] viridis_0.6.5           fastDummies_1.7.3       MASS_7.3-61             DelayedArray_0.30.1     tools_4.4.2             lmtest_0.9-40          
 [55] httpuv_1.6.15           future.apply_1.11.2     goftest_1.2-3           glue_1.7.0              nlme_3.1-165            promises_1.3.0         
 [61] grid_4.4.2              Rtsne_0.17              cluster_2.1.6           reshape2_1.4.4          generics_0.1.3          gtable_0.3.5           
 [67] spatstat.data_3.1-2     tidyr_1.3.1             data.table_1.15.4       sp_2.1-4                utf8_1.2.4              XVector_0.44.0         
 [73] spatstat.geom_3.3-2     RcppAnnoy_0.0.22        ggrepel_0.9.6           RANN_2.6.1              pillar_1.9.0            stringr_1.5.1          
 [79] limma_3.60.4            spam_2.10-0             RcppHNSW_0.6.0          later_1.3.2             splines_4.4.2           dplyr_1.1.4            
 [85] lattice_0.22-5          survival_3.7-0          deldir_2.0-4            tidyselect_1.2.1        locfit_1.5-9.11         miniUI_0.1.1.1         
 [91] pbapply_1.7-2           knitr_1.48              gridExtra_2.3           edgeR_4.2.2             scattermore_1.2         xfun_0.46              
 [97] statmod_1.5.0           stringi_1.8.4           UCSC.utils_1.0.0        lazyeval_0.2.2          codetools_0.2-19        tibble_3.2.1           
[103] cli_3.6.3               uwot_0.2.2              xtable_1.8-4            reticulate_1.38.0       munsell_0.5.1           Rcpp_1.0.13            
[109] globals_0.16.3          spatstat.random_3.3-1   png_0.1-8               spatstat.univar_3.0-0   parallel_4.4.2          ggplot2_3.5.1          
[115] dotCall64_1.1-1         listenv_0.9.1           viridisLite_0.4.2       scales_1.3.0            ggridges_0.5.6          crayon_1.5.3           
[121] SeuratObject_5.0.2      leiden_0.4.3.1          purrr_1.0.4             rlang_1.1.4             cowplot_1.1.3  

[Alexis-Varin]

Hi @Crml23,
Upon looking at the code in fitGAM(), it basically loops over each gene and uses three different functions:
when parallel = TRUE, it uses BiocParallel::bplapply() which is the one using BPPARAM
when parallel = FALSE, but verbose = TRUE, it uses pbapply::bplapply() which in theory should be sequential (like classic lapply()) but only add a progress bar
when parallel AND verbose = FALSE, it uses lapply()

So I would suggest to test with verbose = F see if it still uses all cores. If that's still the case, it could only be from the function that fits the GAM mgcv::gam() internally called for each gene inside the loop, fitGAM() offers control through its parameter control = mgcv::gam.control(), documentation here. the two parameters that matter here are nthreads and ncv.threads which are by default 1, but who knows if on Linux something breaks in the parallelization of the C code. I personally struggled A LOT with nested parallelization (48 cores each using 48 threads and completely overloading the cluster computer) before using a very custom parallelization script with batchtools on our Slurm cluster so I'm guessing that's where something goes wrong on Linux.