GPU queues and costcodes

1. Resources

CellBender runs on a GPU. Choose the LSF queue with --gpuqueue:

--gpuqueue	Costcode	Notes
gpu (default)	Not allowed	The shared farm22 GPU queue
cub22-inference	Required	Billed to the costcode
tiger22-inference	Required	Billed to the costcode

nextflow run cellgeni/nf-cellbender -r main --samples samples.csv \
    --gpuqueue cub22-inference --costcode s1234

The costcode is passed to LSF as -G <costcode>. It is trimmed and lowercased, so S1234 and ` s1234 ` both become s1234.

A costcode is rejected on the gpu queue: it usually means --gpuqueue was forgotten and the jobs would go to the wrong queue. A cub or tiger queue without a costcode is rejected as well.

Resources

Each CellBender job asks for:

Resource	Value	On retry
CPUs	2	Unchanged
Memory	30 GB	30 GB × attempt
GPU	1, shared mode, 8,000 MB GPU memory	Unchanged

These are set in configs/cellbender_removebackground.config. To change them for one run, give your own config with -c:

// my.config
process {
  withName: '.*CELLBENDER_REMOVEBACKGROUND_V0[23]' {
    memory = { 60.GB * task.attempt }
  }
}

nextflow run cellgeni/nf-cellbender -r main --samples samples.csv -c my.config

The other steps run on CPU queues: the iRODS download on transfer (at most 5 at a time), QC and MultiQC on normal.