You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Calibration: should a first run's quick profile later upgrade to standard? #591
SEP 46 §12 item 1, left open at Final (#582). A first native run with no profile calibrates at the quick level (about 20 s) and keeps that row; a standard row (about 90 s) comes only from solprop calibrate or a scan, which wants one for its workers.
Should a quick row later upgrade itself to standard, for example in a background child after the run ends, or stay as it is until the user calibrates?
What is known
On an Intel Xeon Platinum 8581C (Emerald Rapids) the quick level admitted only the medium width: tracking set 0.999 (ctc 0.90), against 0.989 (ctc 0.87, gnlse_raman_shock 0.94) at thorough (docs/dev/benchmarks.md, "Machine calibration against the built-in table").
On the other machines measured both levels kept the table.
A background calibration would contend with the user's next run, and a busy host makes the calibration inconclusive.
Done when
A decision, measured on a machine where quick and standard differ, recorded in docs/guide/performance/calibration.md.
Activity
SEP 46: Final