[Chicken-users] Re: nursery logic

[Brandon J. Van Every]

Brandon J. Van Every wrote:
> I have decided that this test is frail.  On my machine, whether I end 
> up with a nursery that's 256K or 512K in size is pretty random.  The 
> testing is rather inaccurate; nsample -s:128k varies between 380 ms 
> and 480 ms.  Given the multitasking nature of build machines, I would 
> say 3 testing samples aren't nearly enough to get a good average for 
> the nursery size.  I'm going to try 10 samples.
>
> I've decided to prefer smaller nurseries, if performance is equal.  
> Partly due to lack of feedback (Felix?), partly because I don't trust 
> the nsample benchmark.

Even when taking 100 samples, the results are somewhat random.  I've 
seen 48K beat 16K by a hair.  The average results are all within 3 ms no 
matter what the nursery size, leading me to believe that the nursery 
size simply doesn't matter for performance.  Or else the benchmark 
methodology is inadequate; for instance, too much process invocation 
overhead.

I've decided to implement a NOISE_THRESHOLD.  If the average samples 
aren't 5% better than the previous measurement, they are blown off as 
noise.  The 1st measurement is the DEFAULT_TARGET_STACK_SIZE, so to do 
something other than the default, a 5% improvement will be required.  
Using a mere 10 sample runs, which is not entirely adequate and has some 
variation to it, I can't even observe a 2% performance improvement 
consistently.

I am tempted to say the nursery is a waste of time, but perhaps other 
people will have other experiences on different machines.  I'd like to 
hear about people actually getting a 5% improvement.

Anyways the nursery support is in Darcs now.


Cheers,
Brandon Van Every