// nf-core/configs: Medical College of Wisconsin (MCW) RCC profile
//
// Cluster: MCW Research Computing Center (RCC)
// Scheduler: SLURM
// Container engine: Apptainer 1.4.x (drop-in replacement for Singularity)
//
// Partition characteristics observed on this cluster:
//   normal (default) — 60 × (48 CPU / 360 GB / 7 d) — MaxMemPerCPU 7680 MB
//   bigmem           —  2 × (48 CPU / 1.5 TB / 7 d) — MaxMemPerCPU 31500 MB
//   gpu              — 11 nodes, mixed V100 / A40 / L40S, 7 d max
//
// Jobs that would exceed the normal partition's 7.5 GB-per-core cap, or
// its 360 GB-per-node cap, are routed to `bigmem` automatically. GPU
// jobs (label `process_gpu`) are routed to `gpu` and request one device.
//
// Filesystem note: `/scratch` is mounted on compute nodes; `/group` is
// NOT (login-node only). All inputs, outputs, work dirs, samplesheets,
// and the apptainer cache MUST live under `/scratch/g/<user>/` or they
// will appear empty to running tasks. The container bind list below
// intentionally omits `/group`.
 
params {
    config_profile_description    = 'Medical College of Wisconsin (MCW) Research Computing Center cluster profile provided by nf-core/configs.'
    config_profile_contact        = 'Nick Semenkovich'
    config_profile_contact_github = '@semenko'
    config_profile_contact_email  = 'nsemenkovich@mcw.edu'
    config_profile_url            = 'https://www.mcw.edu/departments/clinical-and-translational-science-institute/research-services/research-computing'
}
 
// NOTE: Apptainer is gated behind a module on this cluster, so users
// MUST `module load apptainer/1.4.1` (or any 1.4.x release) AND
// `export APPTAINER_CACHEDIR=/scratch/g/$USER/.apptainer/cache` (plus
// APPTAINER_TMPDIR) in their submission script before invoking
// `nextflow run`. The driver-side image pull and OCI→SIF conversion
// happen before any `process.beforeScript` or `env { }` scope takes
// effect, and home directories on this cluster are too small to hold
// a fresh apptainer working set. See docs/mcw_rcc.md.
apptainer {
    enabled    = true
    autoMounts = true
    runOptions = '-B /scratch,/hpc'
    // Persistent image store: one shared `.img` cache per user under
    // /scratch. Equivalent to setting NXF_APPTAINER_CACHEDIR; doing it
    // here means users don't have to export it. Apptainer's own
    // pull-time cache (APPTAINER_CACHEDIR / APPTAINER_TMPDIR) still
    // needs to be exported in the driver script — see docs/mcw_rcc.md.
    cacheDir   = "/scratch/g/${System.getenv('USER')}/nf-apptainer-cache"
}
 
// Task-side safety net: any container engine operation that happens
// inside a process picks up these paths. Users who forget to export
// them in the driver script will still see the driver-side pull
// fail, but task re-pulls and any in-task apptainer use are routed
// to scratch automatically.
env {
    APPTAINER_CACHEDIR = "/scratch/g/${System.getenv('USER')}/.apptainer/cache"
    APPTAINER_TMPDIR   = "/scratch/g/${System.getenv('USER')}/.apptainer/tmp"
}
 
process {
    executor = 'slurm'
 
    // Single-node ceilings: bigmem nodes are 48 CPU / 1.5 TB; max walltime 7 days.
    resourceLimits = [
        memory: 1500.GB,
        cpus  : 48,
        time  : 168.h,
    ]
 
    // Apptainer replaces the old `singularity` module on this cluster.
    beforeScript = 'module load apptainer/1.4.1'
 
    cache = 'lenient'
 
    queue = {
        def mem_per_cpu_mb = task.memory ? (task.memory.toMega() / task.cpus) : 0
        if ((task.memory && task.memory > 360.GB) || mem_per_cpu_mb > 7680) {
            return 'bigmem'
        }
        return 'normal'
    }
 
    withLabel: 'process_gpu' {
        queue            = 'gpu'
        clusterOptions   = '--gres=gpu:1'
        containerOptions = '--nv'
    }
}
 
executor {
    queueSize       = 100
    submitRateLimit = '10 sec'
    pollInterval    = '30 sec'
}