[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Gcl-devel] Re: surprising slowdown

From: Camm Maguire
Subject: [Gcl-devel] Re: surprising slowdown
Date: 23 Jan 2007 17:30:52 -0500
User-agent: Gnus/5.09 (Gnus v5.9.0) Emacs/21.2

Hi Matt!

Yes this is quite surprising.  I'm running through your build
instructions now on black.  The most useful profiling information
comes from the profiling build we discussed some time ago -- I've at
least been supporting this in the Debian packages as this build method
is required still on some machines.  Do you still maintain the gprof
file in the current acl2 distribution?  If so, will it work if I
substitute loading this file for your previous build instructions?

Had an idea regarding more flexible profiling without requiring a
separate image, though the available information is more limited.  Run
with fast links off, and supply a little C function to hash the
caller, callee, and time spent, then write some diagnostic routine to
put the information together in a graph.  The ingenious fast-link
mechanism makes this possible without having to run slower when not
profiling.  As Bob Boyer noted earlier, making the call through a
pointer (which provides this flexibility) costs a mere instruction per
call on x86.

Will ry to get something like the above into 2.7 soon.

Take care,

Matt Kaufmann <address@hidden> writes:

> Hi, Camm --
> I found a truly surprising performance issue, running GCL 2.6.7 on
> 32-bit linux.  It seems that by moving a single function definition
> from an interpreted file to a compiled file, I would get a 2% or 3%
> slowdown in the ACL2 regression suite.  Several tests consistently
> point to this, and I wonder what's going on, and I think you might
> wonder too.  Here is a typical result.
> faster (defined in interpreted file):
> 4882.245u 117.395s 1:25:12.71 97.7%   0+0k 0+0io 0pf+0w
> slower (defined in compiled file):
> 5040.126u 116.955s 1:27:43.68 97.9%   0+0k 0+0io 0pf+0w
> I've put a tarfile camm.tgz in my home directory on the UT CS file
> system, since you have an account there.  I'll email it to you if you
> prefer (about 4.4M).  When you extract, the README file explains what
> is going on in detail and shows you how to re-create the results.
> Thanks --
> -- Matt

Camm Maguire                                            address@hidden
"The earth is but one country, and mankind its citizens."  --  Baha'u'llah

reply via email to

[Prev in Thread] Current Thread [Next in Thread]