User Manual
Complete guide to using CAC computing resources
Memory Requests — the basics
From 1 September 2026, memory requests are enforced on GRAMMAR
Until now --mem was recorded but not applied. After
the switch, a job that exceeds its request is terminated with OUT_OF_MEMORY — instead of taking the whole
node down and killing everyone else's jobs with it.
Which directive to use
| Directive | Meaning | Use when |
|---|---|---|
| --mem-per-cpu=4G | Memory per allocated core | Core count varies (array, MPI) |
| --mem=64G | Total per node, all processes combined | Fixed core count, single process |
Compute nodes have 64 cores and 503 GB — about 8 GB per core.
That ratio is the balance point: ask for more memory per core and your job blocks cores it never uses,
which lengthens your own queue time. Omit both directives and 8 GB per core is assigned automatically.
(The debug partition node has 251 GB, so its balance
point is about 4 GB per core.)
--mem is the total for the entire node,
not the amount for one process or worker. If your script launches N workers, the request must cover
all N. Misreading this is the single most common cause of node failures on our clusters.
Checking what your job actually used — and the rest of the manual — requires a login. Two entry points that do not:
- FAQ › Job Submission — how much to request, what to do when a job is killed with
OUT_OF_MEMORY - On the cluster:
sacct -j <jobid> --format=JobID,ReqMem,MaxRSS,Statefor a finished job. Request about 1.3–1.5× the peak it reports.
Login Required
This manual is only available to registered users.
Please log in to access the documentation.
Don't have an account? Apply for access