Amdahl’s Law.

A part that cannot be sped up limits the gain from adding more workers.

Interactive experimentintuitiveField note ·

How many processors are worth adding?

A part that cannot be sped up limits the gain from adding more workers.

ILLUSTRATIVE MODEL

Parallelizable work80%

10%99%

Processors8

132

PLAY THE IDEAMake a choice

Move a control above. Compared with the starting setup, will speedup rise, fall or stay the same?

Choose a move to see what happens.

Speedup

Adding workers gives diminishing returns because the serial part remains.

Coordination overhead and changing workload are excluded, so real gains can be smaller.

Preparing the experiment…

THE IDEA

Speeding up one part of a job leaves the untouched part as a bottleneck.

Speedup with n processors is 1/[(1−p)+p/n], where p is the parallelizable share. The plot shows speedup for 1–32 processors at the selected share.

Estimate whether parallelizing a slow data task will justify more servers.

WHERE IT BREAKS

Coordination overhead and changing workload are excluded, so real gains can be smaller.