A simulation has a life before and after its compute phase. Data preparation, transfers, checkpoints and analysis all contribute to elapsed research time.
A large sequential-bandwidth figure says little about a workload creating many small files or sharing storage with several jobs. The access pattern and concurrency have to match the test.
Build a simple timeline from staging to validated result. Assign capacity and retention ownership, then test a restart. Archive and deletion policies should follow the research and data requirements.
Use it in the next discussion
- Time stage-in, run, checkpoint and stage-out separately.
- Include file counts, concurrency and recovery.
- Check whether the fastest compute configuration moves the bottleneck elsewhere.
PRIMARY SOURCES · CHECKED 9 NOVEMBER 2026
NERSC: Roofline performance model →Living documentation; checked 2026-11-09 · Relates arithmetic intensity to compute and memory limits; measured kernels still matter.
NERSC: job best practices →Living documentation; checked 2026-11-09 · Facility-specific operational guidance; use local site policies for actual jobs.