Strong scaling holds the total problem size fixed. Weak scaling increases total work while keeping work per processor fixed. The two experiments answer different questions.
Before selecting a system size, capture a baseline run and repeat the same workload at several resource counts. Record runtime variance, communication, memory pressure and I/O rather than reporting only the best run.
The Scaling Lab uses a transparent Amdahl-style model to illustrate a serial fraction. Its overhead input is an explicit additional assumption at the selected node count. It cannot predict a specific machine without measurement.
Use it in the next discussion
- Keep input, software and numerical tolerance constant.
- Show time to solution and node-hours together.
- Treat a model curve as a question to test, not a benchmark result.
PRIMARY SOURCES · CHECKED 9 NOVEMBER 2026
NERSC: parallelism →Living documentation; checked 2026-11-09 · Strong scaling fixes total problem size; weak scaling fixes work per processor.
NERSC: Roofline performance model →Living documentation; checked 2026-11-09 · Relates arithmetic intensity to compute and memory limits; measured kernels still matter.