CISL has implemented the changes that were announced in January to account for the use of GPU resources on the Casper cluster. The charging algorithm update for jobs on shared CPU nodes on both Casper and Cheyenne also has been implemented.
Projects that have used Casper GPU resources in the past year have all now received an initial GPU-hours allocation based on that previous use. Charges will be based on a job’s walltime. To check the status of those allocations, go to the Systems Accounting Manager
My Projects page (login required). Submit any requests for additions through the
NCAR Research Computing help desk.
The revised shared node charging algorithm is based on elapsed walltime rather than Unix CPU time, to be consistent with GPU-hour accounting and with the planned accounting for Derecho use when that system becomes available.
The shared-node algorithm change will have no major effect on allocation use for many workflows, but will result in increased charges for jobs that leave compute cores idle and unavailable to other users. To avoid unnecessary charges and to release idle resources that others could be using, free up those resources (for example, by terminating JupyterHub batch servers and closing Dask clusters).
