[m-dev.] Timings vs SICStus

[Zoltan Somogyi]

> We're trying to do some timing comparisons between HAL programs compiled to 
> Mercury, and SICStus programs.  Is there a paper to go along with the table 
> of benchmarking results on the web site?

Yes, the execution algorithm paper in JLP; details ond postscript on the web.

> Mainly I just want to verify what options you 
> used with SICStus,

I don't remember using any options beyond fastcode and compactcode.

You can see the whole experimental setup in /home/mercury1/zs/bench.tar.gz.

> and why you seem (from what I can tell from the 
> benchmarking code) to have ignored the time SICStus spends garbage 
> collecting, stack shifting, making system calls, etc.

We measured these microbenchmarks by executing them lots of times
in a failure driven loop, subtracting the time at start from the time at
end, and dividing by the number of repetitions. Each execution is so quick
(I think one was 6 microseconds) that SICStus spends *no* time doing gc
or stack shifting or making system calls, and the failure driven loop
recovers all system resources between benchmark invocations.

Zoltan.