Instructor Notes
Timing
| Episode | Teaching | Exercises |
|---|---|---|
| How to evaluate a run | 15 | 5 |
| What limits a run | 20 | 10 |
| Strong and weak scaling | 15 | 20 |
| The roofline | 25 | 20 |
| Energy | 10 | 10 |
| Slurm resources | 10 | 15 |
What to cut
The energy episode is the first to cut. EAR and eacct
exist on Snellius. The course tab says to skip them.
Numbers you may quote
The 8-core efficiency 0.625 comes from the worked series 8.0 seconds
on 1 core and 1.6 seconds on 8 cores. Show 8.0 / (8 * 1.6).
Keep that series separate from any AXPY times the room records.
The roofline example is throughput 6.7257 and bandwidth 2.6767, so AI = 6.7257 / 2.6767 = 2.5 FLOP/byte. Do not substitute a different arithmetic intensity. The deck does not give a numeric ridge. Learners compare 2.5 with the ridge of the machine they measured.
The AXPY binary is in the hybrid-jobs lesson. Do not paste a second
axpy.c into this lesson.
Sam
Sam wants a number for whether more cores helped. Give Sam speedup
T(1) / T(N) and efficiency T(1) / (N * T(N)).
A shorter real time means the run finished sooner.
Efficiency is how much of the added cores showed up in that time.
Tabs
The course cluster is a SURF Research Cloud workspace, and the course
tab requests --partition=rome. Counters on the course
cluster are perf and PAPI, not uProf. Billing units, if
anyone asks, are the accounting tab. Do not put a second copy of that
formula on the board.