Re: [Chicken-users] Nursery sizing considered stupid

[John Cowan]

felix winkelmann scripsit:

> cd to benchmarks, compile one (say "csc -Ob dynamic"), run it with
> nursery settings -:s32k, -:s300k and the default, multiple times.

Okay.  I also added 128K, my proposed default; the default on my current
build is 256K.  I ran 100 separate tests using a shell for-loop for
each size.  Here's what I got:

Nursery size            Mean            Standard deviation
32K                     0.208s          0.011s
128K                    0.159s          0.014s
256K                    0.154s          0.009s
300K                    0.151s          0.010s

This is on Cygwin using a machine as idle as I could reasonably make it
short of unplugging the network cable.

Obviously 32K is a bad idea, and I will leave it up to those with more
statistical knowledge than myself whether these results are statistically
significant (what stat I ever knew dissolved in a cloud of rust years
ago), but eyeballing them I'd say there's not much to choose from in
any of the last three.

> Then tell me the nursery size doesn't make a difference. It does, at
> least on Linux machines.

My claim has never been that the nursery size as such makes no difference,
but that the method for determining the best nursery size doesn't actually
do so -- particularly under realistic conditions for a build.  It weeds
out the worst sizes but otherwise picks one more or less at random.
Therefore, all it does is slow down the build.

-- 
Some people open all the Windows;       John Cowan
wise wives welcome the spring           address@hidden
by moving the Unix.                     http://www.ccil.org/~cowan
  --ad for Unix Book Units (U.K.)
        (see http://cm.bell-labs.com/cm/cs/who/dmr/unix3image.gif)