Glossary

Glossary

---

wall-clock time

The real line from time -p. The seconds you waited.

strong scaling

The problem size stays fixed and the number of processors grows. Speedup is T(1) / T(N).

weak scaling

The problem size grows with the number of processors, so each processor keeps about the same work.

speedup

T_s / T_N. T_s is the serial time. T_N is the time on N processors.

efficiency

T_s / (N * T_N). The worked 8-core value in this lesson is 0.625.

Amdahl’s law

Strong-scaling speedup is 1 / (s + p/N). s is the serial share of the time, p is the parallel share, N is the number of processes.

Gustafson’s law

Weak-scaling speedup is s + p*N. The parallel share grows with the resources. The serial share stays fixed as the problem grows.

arithmetic intensity

Floating-point operations per byte moved. The unit is FLOP/byte.

ridge point

The arithmetic intensity where the bandwidth roof meets the peak roof.

roofline

A plot of performance against arithmetic intensity, under the bandwidth roof and the peak roof.

memory-bound

Left of the ridge. The kernel waits on bytes moving.

compute-bound

Right of the ridge. The kernel waits on arithmetic peak.

FLOP

One floating-point operation. An add is 1 FLOP. A multiply is 1 FLOP.

EAR

Energy Aware Runtime on Snellius. --ear=on and --ear-policy=monitoring, then eacct -j JOBID.

Tools from the deck

---

Tool	Cost	Description
Linaro DDT	Non-free	Full featured graphical, parallel debugger
HPCToolkit	Free	Integrated suite of tools for parallel program performance analysis
Intel One API	Free, with conditions	Stack of performance analysis and debugging tools (MPI/OpenMP/SIMD)
Valgrind	Free	Memory errors debugging tool
TotalView	Non-free	Full featured graphical, parallel debugger
Vampir	Non-free	Full featured trace visualizer for parallel program OTF trace files
memP	Free	Lightweight memory profiling tool
mpiP	Free	Lightweight MPI profiling tool
PAPI	Free	A standardized and portable API for accessing performance counter hardware
likwid	Free	A tool to measure hardware performance counters
Tau	Free	Full featured parallel program performance analysis toolkit
Extrae	Free	MPI/OpenMP profiler
Scalasca	Free	MPI+OpenMP analysis with Score-P (Score-P/9.2-gompi-2025a)
Darshan	Free	IO profiler
Nsight	Free	GPU: nsys (Nsight-Systems/2025.4.1), ncu (CUDA/12.8.0)
Linaro MAP	Non-free	Performance analysis tool for MPI+OpenMP
uProf	Free	AMD CPU counters, sampling, and roofline analysis
gprof	Free	Standard unix/linux profiling utility

The course cluster tab in the roofline episode uses perf and PAPI. uProf is the Snellius example in the deck.