Parallel performance workbench
Find the workers
worth adding.
Compare the same workload across worker counts. Find out where more compute buys meaningful time and which measured configuration fits your deadline.
01 / Measured runs
Paste your timings
02 / Decision
Speed versus compute
Compare at least two worker counts to see where runtime and worker-seconds diverge.
- Fastest measured
- Least compute
| Workers | Median | Speedup | Efficiency | Worker-s | Cost |
|---|
Measurement notes
Hold the workload constant.
Use the same input, worker type, and timing boundaries for every run. The tool groups runs by worker count and takes the median. It keeps the fastest and slowest run in the export so you can see variability.
Speedup is the one-worker median divided by the selected median. Efficiency is speedup divided by worker count. Worker-seconds are median elapsed time multiplied by the number of workers. The tool selects the smallest measured worker-seconds among configurations whose median meets the deadline. It does not predict future runtimes.
Optional cost multiplies worker-hours by your rate. This assumes workers are reserved for the full measured duration at a common rate. Add storage, billing minimums, and other charges separately.
For fixed-workload scaling context, see Gene Amdahl’s original discussion of parallel speedup.