Previous | Next --- Slide 9 of 44
Back to Lecture Thumbnails
sgbowen

I'm not sure I understand exactly what this slide is saying about pitfalls, so I'll rephrase it just to be sure...

Fixed problem size speedup analysis means: keeping the size of the problem (N) the same, try running the problem on a different number of cores and use those results to determine how well your parallel algorithm works.

As a result, you'd see only one of the lines in the graph, and you might erroneously conclude there's something wrong with your parallelization.

Presumably the solution is to try several different problem sizes and see if you note some difference in the slopes of the lines.

rramo

@sgbowen With the reason for this being that we must account for communication-to-computation ratios. Meaning that for small problem sizes, the overhead of communication shadows any speedup gained from the parallelism.