Amdahl’s Law describes the maximum speedup that can be achieved through parallelisation: it is bounded by the serial fraction. However, in 1988, John Gustafson countered that if the problem grows in proportion to the resources available, the speedup remains almost linear. This calculation works out in high performance computing. However, for load-driven web systems, it is ineffective because Gustafson’s optimism hinges on a constant serial fraction, which grows under load instead of shrinking.
The latest instalment in the Scalability Series shows when Amdahl’s law and Gustafson’s law are the right model, and explains why the two are not competitors but answer different questions.