Checking Resource Usage & Billing

Rockfish charges jobs in core-hours. A core-hour is one CPU-core used for one hour — or the equivalent, when GPUs or very large memory requests are involved.

Core-only partitions

The classic formula is:

Core-hours =   Allocated CPU cores   ×   Wall-time (h)

Examples:

1-hour job using 16 cores → 16 core-h
6-hour job using 4 cores → 24 core-h

To inspect the actual billing for a finished/active job:

sacct -j <jobid> --format=JobID,User,Partition,State,Elapsed,CPUTime,NCPUS,AllocTRES

Viewing Allocation Usage with `sbalance`

To view usage for a Slurm Allocation over the course of a billing quarter, use the sbalance tool. For example:

sbalance

The time of this report is: Fri May 09 11:42:39 2025

Start Date: 2025-04-01T00:00:00 End Date: 2025-06-30T00:00:00
                                          Used          Account
               Account        User   (core-hours) /    Allocated )
====================================================================
              pi               870000.0 /      900000.0
....................................................................
                               user01            250000.0
                               user02             50000.0
                               user03            100000.0
                               user04             75000.0
                               user05             40000.0
                               user06             30000.0
                               user07             20000.0
                               user08             15000.0
                               user09             10000.0
                               user10             25000.0
                               user11             35000.0
                               user12             50000.0
                               user13             60000.0
                               user14             10000.0
                               user15             10000.0

Use --usage to include a usage percentage column:

sbalance --usage

This will show the same table, but with each user’s percent of the allocation.

You can also use --hide-zero to omit users who have not yet used any core-hours.

GPU partitions

GPU nodes are billed at a rate of: (GPUs requested * Billed Cores / Node) * Walltime.

The ratio of “Billed Cores / GPU” comes from the number of available cores and GPUs on the node. For example, in the A100 partition, each node has 4 GPUs and 48 CPU cores, so the ratio is 12.

Partition	GPUs / node	Billed cores / GPU
`v100`	4 × V100	12
`a100`	4 × A100	12
`l40s`	8 × L40S	8
`ica100`	4 × A100	16
`mig_class`	12 × A100-MIG (2g.20gb)	5 (per slice)

Billing formula

Core-hours = ( GPUs × billed-cores/GPU ) × Wall-time(h)

Example for a100 partition:

2xA100 GPUs for 3 hours → (2 × 12) × 3 = 72 core-hours

Bigmem partition

On bigmem jobs are billed on the largest of (CPU cores) or (memory/32 GiB), because each core is paired with 32 GiB of RAM:

Effective cores = max( requested CPUs ,
                       requested memory (GiB) / 32 )

Core-hours      = Effective cores × Wall-time(h)

Example for bigmem partition:

400 GiB Memory for 4 hours → (400/32) * 4 = 50 core-hours

Viewing Historical Usage and Efficiency

For more on viewing efficiency and job usage statistics, refer to the page:

Viewing Job Status & Efficiency

This includes guidance on using:

sacct for historical usage
seff and reportseff for job efficiency
jobstats for GPU and memory metrics