The HLL Performance Gap

Patrick R. Michaud pmichaud at pobox.com
Wed Jul 22 00:57:00 UTC 2009


On Tue, Jul 21, 2009 at 04:31:20PM -0700, Geoffrey Broadwell wrote:
> THE ANALYSIS:
> 
> A few weeks ago, I was able to spend some time trying to track down why
> this huge performance gap existed, at least for my code running in
> Rakudo.  I suspect that other HLLs face similar problems, and I'd love
> to see what the rest of you have found.
> 
> In any case, I found a number of issues, summarized in this list:
> 
> 1. Every Perl 6 scope becomes a PIR sub.
> 2. Rakudo scopes are extra-heavy.
> 3. PCT has no optimization passes.

All of these are valid issues, but personally I find that they are
of *far* less importance than

4.  Parrot subroutine calls are far too slow
    4a.  There is too much overhead in parameter passing

So while I agree that we can get started on things like PCT optimizations,
of far greater importance to HLL performance will be resolving
Parrot's calling conventions and in speeding up the basic mechanisms
for invoking and returning from Parrot subroutines.

Phrased another way -- based on the measurements I've done, the
speed overhead that currently exists in Rakudo arising from
variable setup and takedown is far outweighed by the basic
costs of performing subroutine calls.  While we may be able to
eliminate some subroutine scopes (and I already have designs
for doing this), the number of cases where this will in fact
be possible is quite small.  Based on my experience in developing
on Parrot, I suspect the fixes and optimizations you're describing 
for Rakudo are in fact more of the "iterating on fixes that each 
gain a few percent" variety than the "order of magnitude improvements".

Again, I'm not saying that we shouldn't start looking at PCT
improvements.  But I am saying that it's far more important
that we stabilize Parrot's calling conventions early in the
process -- not only for the performance improvements to be
gained from doing so, but also because our ability to
measure and adequately optimize PCT directly depends on the
underlying stratum being stable and efficient.

Pm


More information about the parrot-dev mailing list