Gustafson’s Law
Ideal scaled-workload speedup when more processors allow a larger parallel task in the same elapsed time. Here s is the serial fraction of the parallel execution time, not the single-processor fraction used in Amdahl’s law.
Formula
Live Calculator
MembersVariables
Example
With P = 16 and s = 0.10: S = 16 − 0.10 × 15 = 14.5. This is ideal scaled speedup, not a guarantee that a fixed task will run 14.5 times faster.
How it works
More processors can let a team tackle a larger simulation instead of merely finishing the original simulation sooner. Gustafson’s law models that scaled-workload perspective. Normalise the parallel run time to one: serial work takes s and parallel work takes 1 − s. The same enlarged workload on one processor would take s + P(1 − s), which simplifies to P − s(P − 1).
When to use it
Estimate ideal speedup when the parallel part of a workload grows with processor count. It assumes the parallel work can be distributed ideally and omits communication, synchronisation and load-balancing costs.
Worked example, step by step
Step 1: Use P = 16 processors. Suppose serial work is s = 0.10 of the parallel run time.
Step 2: Compute S = P − s(P − 1).
Step 3: S = 16 − 0.10 × (16 − 1).
Step 4: S = 16 − 1.5 = 14.5.
Cross-check: 0.10 + 16 × 0.90 = 14.5.
Interpretation: under the model, a one-processor run of the same enlarged task takes 14.5 times as long as the parallel run. This does not promise that a fixed-size task becomes 14.5 times faster.
Common mistakes to avoid
- Using the single-processor serial fraction from Amdahl’s law as if it were Gustafson’s parallel-time fraction.
- Claiming scaled-workload speedup for an unchanged workload.
- Ignoring communication costs and uneven work distribution.
- Treating an ideal estimate as a measured benchmark.
Did You Know?
Gustafson’s law and Amdahl’s law answer different scaling questions: growing the problem versus keeping the workload fixed.
Share this formula
More in Computer Science
View allHick–Hyman Law (Equal Choices)
IntermediateModels choice reaction time for n equally likely alternatives (n ≥ 1). This information-theoretic form assumes comparable, practised stimulus–response choices; it is not a universal law of menu browsing.
Fitts’s Law (Shannon Form)
IntermediateAn empirical model of aimed movement time: farther or narrower targets generally take longer to reach. Use D ≥ 0, W > 0 and fitted coefficients for the device and task.
Shannon Entropy
AdvancedAverage information content of a random source (bits).
Master Theorem (form)
AdvancedAnalyzes running time of divide-and-conquer algorithms.
Hamming Distance
AdvancedNumber of positions at which two strings differ.
Bits to Represent n
BasicMinimum number of bits to represent n distinct values.