When we parallelize tasks and measure the perforamnce,two common measures of parallel scaling are

Strong Scaling: Time for fixed problem size as number of processor is increased

Weak Scaling: Time for fixed computational work per processor as problem size is increased

Some documents mention that the weak scaling is easier to achieve (most of the cases) than the strong scaling.

Is this correct? if then, why?

  • 1
    $\begingroup$ Simple explanation: communication and data dependencies. Increasing the problem size tends to increase the amount of isolated work available relative to communication. For example, some physical modeling only communicates between nearest neighbors at the end of each time step; increasing the number of cells simply increases parallelism, having each processor perform a fraction of the time steps work for a cell would typically mean that data within a cell would have to be communicated between processors and any data dependencies could force waiting for results. $\endgroup$ – Paul A. Clayton Nov 27 '16 at 9:24
  • 1
    $\begingroup$ For a purely theoretical approach, strong scaling is related to Amdahl's law (at fixed size you are limited by the part that cannot be made parallel), while weak scaling is better described by Gustafson's law (if the increased problem size only affects the parallel part of the program, you can handle bigger problems with more parallelism) $\endgroup$ – Gabriel Gouvine Nov 27 '16 at 12:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.