Are you graphing just GC time, or overall execution time here? It would be better to use the latter, because GC behaviour can impact locality and change the mutator time too.

I’d also be very interested to see similar graphs showing the memory footprint.

a) A linear cost for heap size, which causes time to rise for huge -H values. malloc(1G) doesn’t seem to cost any measurable amount on my computer, does ghc do something to the allocated memory, or what?

b) an inverse cost for allocation area size, making very small -A values quite expensive, and responsible for the “ski-jump” shape for low -H

c) -A cost becomes mostly irrelevant when -H is big enough – perhaps to fit the working set – eliminating the ski jump.

GHC defaults are -H 0 and -A 512K. These are small benchmarks, and in general, I don’t think we can expect -H to be big enough to avoid the ski jump.

It’d be interesting to see how the other benchmarks fare, and also how things work out on other CPUs with different cache sizes etc. Perhaps information from -s output (number of collections etc) could be incorporated? Anyway, great work, Don!