Paritian

Developers

Gustafson's Law Calculator (Scaled Speedup)

Speedup when the problem grows with the number of processors, by Gustafson's law, set beside Amdahl's fixed-size speedup.

Results

Scaled speedup (Gustafson) 60.85 ×
Parallel efficiency 95.1 %
Fixed-size speedup (Amdahl), for comparison 15.42 ×
Amdahl's ceiling with infinite processors 20.0 ×

What this tool does

Amdahl's law is gloomy: with 5 % of the work serial, no number of cores can make a job more than 20 times faster. John Gustafson pointed out in 1988 that people rarely run the same small job on a bigger machine; they run a bigger job in the same time, and then the serial part shrinks in relative terms. With 64 cores and 5 % serial time, the scaled speedup is about 61 instead of Amdahl's 15. That is weak scaling, the model behind supercomputer benchmarks; Amdahl describes strong scaling, which is what you get when you only want the same answer faster. Here the serial share is measured on the parallel machine, as Gustafson defined it.

Formula

S = N − s (N − 1) · Amdahl: S = 1 ÷ (s + (1 − s) ÷ N)

Variables

SymbolMeaningUnit
nNumber of processors or cores—
sSerial share of the work%
GSScaled speedup (Gustafson)×
EFParallel efficiency%
AMFixed-size speedup (Amdahl), for comparison×
CEAmdahl's ceiling with infinite processors×

Worked example

  • Number of processors or cores64
  • Serial share of the work5 %
  • Scaled speedup (Gustafson)60.85 ×
  • Parallel efficiency95.1 %
  • Fixed-size speedup (Amdahl), for comparison15.42 ×
  • Amdahl's ceiling with infinite processors20.0 ×

Limitations

  • The result is an estimate based only on the values you type. Real situations often include factors this calculator does not know about.
  • The formula assumes ideal conditions: no friction losses, no air resistance and no efficiency losses unless you enter them.