Antoine Pitrou wrote:
> Antoine Pitrou <pitrou@free.fr> added the comment:
> 
>> I'm not bothered by the regression in "silent_logging",
>> as it is a micro benchmark with a very short running time.
> 
> I'm not concerned about the micro-benchmark itself but the fact that it
> might hint at a wider problem.

Or it might not.
Micro-benchmarks produce micro-optimisations.
That's why I dislike them.

> Also, I don't get your remark about it running in a short time. Your
> patch AFAICT doesn't need any warm up period to exhibit any
> improvements.

What I mean is that the runtime is so short, no one would notice any
change, so who cares?

> 
>> Reducing the method-cache size from 2**10 to 2**9 allows the working
>> set to fit better in the cache.
>> This fixes the regression in "mako", but makes OO programs that use
>> few objects (such as richards) a bit slower.
> 
> I don't think we should reduce the size of the method cache. 1024 is not
> a very large number, and might even be too small for complex OO
> programs.

"not very large" or "too small", by what metric?

The optimum size (for speed) of the method cache depends on the size of 
hardware data cache, the complexity of the program, and many other factors.
Attempt to reason about it are pretty much futile.
Empiricism is the only way to go.

> 
> I also think that, apart from the dict storage changes, your patch
> should strive not to change any other tunables. Otherwise we're really
> comparing apples to oranges.

If the implementation changes, shouldn't the tunable parameters be retuned?

Cheers,
Mark.