Parallel Computing & Cluster Scaling Estimator
Predict how a job scales with more cores or nodes: fixed-size speedup with Amdahl's law, scaled speedup with Gustafson's law, and parallel efficiency, given the fraction of work that parallelises.
Results
- Amdahl speedup on 64
- 15.42×theoretical
- Parallel efficiency
- 24.1%theoretical
- Gustafson scaled speedup
- 60.8×theoretical
- Maximum speedup (infinite processors)
- 20×theoretical
- Run time on 64
- 6.5 hestimated
| Processors | Amdahl speedup | Efficiency | Gustafson | Run time |
|---|---|---|---|---|
| 1 | 1 | 100% | 1 | 4.2 days |
| 2 | 1.9 | 95.2% | 2 | 2.2 days |
| 4 | 3.48 | 87% | 3.8 | 28.8 h |
| 8 | 5.93 | 74.1% | 7.6 | 16.9 h |
| 16 | 9.14 | 57.1% | 15.3 | 10.9 h |
| 32 | 12.55 | 39.2% | 30.5 | 8 h |
| 64 | 15.42 | 24.1% | 60.8 | 6.5 h |
Communication, load imbalance and memory contention reduce real scaling below Amdahl's prediction.
What it calculates
- Amdahl speedup
- Gustafson scaled speedup
- Parallel efficiency
- Maximum possible speedup
- Scaling table
Worked example
Using the default inputs below; change any input above to recalculate.
| Parallel fraction of the work | 95 % |
|---|---|
| Processors (cores or nodes) | 64 |
| Run time on one processor (optional) | 100 hours |
| Amdahl speedup on 64 | 15.42× | theoretical |
|---|---|---|
| Parallel efficiency | 24.1% | theoretical |
| Gustafson scaled speedup | 60.8× | theoretical |
| Maximum speedup (infinite processors) | 20× | theoretical |
| Run time on 64 | 6.5 h | estimated |
Methodology
Amdahl's law gives the speedup of a fixed-size problem when a fraction p parallelises perfectly. Gustafson's law gives the scaled speedup when the problem grows with the processor count.
Formulas
- Amdahl
- S(n) = 1 ÷ ((1 − p) + p ÷ n)
- Gustafson
- S(n) = n − (1 − p)(n − 1)
- Efficiency
- E = S(n) ÷ n
Assumptions
- The parallel part divides perfectly; overheads are not modelled.
Limitations
- Measure p from profiles or from timings at two processor counts for real predictions.