kantp
commented
8 years ago It seems that the scaling behaviour depends critically on the type of jobs.
In the particle filter tool, we get a very nice linear speedup curve, for the first ~ 10 workers.
I'm trying to replicate this in a simpler benchmark, but currently, I get a decrease in the performance when adding the first slave, and only moderate speedup afterwards. Probably, the initial update steps are too simple, so that the gain by parallelising is outweighed by the administrative overhead.
I'll use an instrumented version of work-queue with the particle filter, in order to get a better understanding of how long the different steps take, in order to design a benchmark with more similar characteristics.
While working on #139, I wrote a small program to measure the performance of the work-queue library. @bitonic suggested to add it as a proper benchmark, which is a good idea.