Fri, 14 Dec 2001 19:21:07 +0000
> On Fri, 14 Dec 2001, Simon Marlow wrote:
> > > I think you should shoot for more like 60 seconds. A test that
> > > runs for 10 seconds today will be back in the nearly-useless
> > > category (~1 second) within 2-3 years, at current progress rates
> > > for hardware.
> > I think 60 seconds would be a bit on the long side - when running tests
> > interactively rather than in the background having tests that run that
> > long would be too painful. 10 or maybe 20 seconds would be better IMO.
> Could there be maybe 3 possibilities - short, medium and long, of maybe 5,
> 20 and 60 seconds, for example? And choose it via the Makefile, or
> something? I don't think one can say "this is the optimal time" - it
> depends on what you're doing...
Sounds like a good idea.
So that's now what I'm doing.
By setting the value of a variable called "mode" in the top level Makefile,
one can switch between three modes.
They are as follows:
fast: Takes around 1 second.
Good for when one is running tests with masses of profiling
going on that make things run very slowly.
normal: Behaves exactly like current NoFib.
(ranges from very fast to very slow)
slow: Takes around 10 seconds.
I might consider an "ultra-slow" mode as well, taking around 60 seconds, but
running tests over and over again until they take around 60 seconds is likely
to be a bit tedious, so that can go rather further down my todo list.
Times are based on how long they take to run on my PIII 750 box when compiled
with the current CVS version of GHC. Not the most objective measure of speed,
but probably a fairly reasonable one.
I'm currently ignoring other changes that we might want made to NoFib programs.