Perl with internal Benchmark/Profiler

[PetaMem]

Hi,

we happen to compile own perl interpreters a lot. Mainly to get up-to-date high-performance Perl Interpreters tailored exactly to our needs. Until now, to check impact of used compiler, compiler flags, perl configuration etc. we simply benchmarked some perl programms of ours.

It would be of great benefit, if there was some pragma or Module belonging to the Perl core, that could give information about the speed of various operations. It is evident, that the result of such benchmarking cannot be some one-dimensional number "this perl interpreter runs with xxx drystones", but rather a table where the speed of list creation, sorting, hash-lookup, iteration etc. would be benchmarked.

Ideally, the output of this would be easily parseable, so we could simply let our compiler run over the weekend and test all the perls with all the relevant flags. This should give us a fairly good overview how the different perls perform. Or didn't you always want to know how a perl-5.9.3 -O3 -msse2 -march=k8 64bit no-multiplicity compares to a 5.8.8 32bit perl -O2 ...

Anyone a hint how to proceed to get an exhaustive set of such benchmark results?

[dave_the_m]

Sounds like you want perlbench.

Note that differences of a few percent are usually noise: the effect of cache alignments etc.

Dave.

[PetaMem]

From what I can say so far: Yes! Strike! That's exactly what I was looking for. Great tool. And I have already seen some very unexpected results. Stay tuned, more to come...

[perrin]

When I did something like this, I found that turning off threads and debugging (not the perl debugger, but debugging perl itself) made a big diference, and using the Intel compiler made a pretty big one too.

[PetaMem]

I can confirm that tendency (from the impressions so far). But my intention is now to get some numbers to get a quantitative statement. Unfortunately, the Intel compiler performs far worse on 64bit Architectures than other options.

[PetaMem]

Could you please elaborate some HowTo use the icc? After extensively using perlbench, I came to the conclusion that gcc 4.x sucks and tried icc. Unfortunately, I'm getting lots of these:

icc -L/usr/local/lib -o miniperl \
    miniperlmain.o opmini.o libperl.a -lnsl -ldl -lm -lcrypt -lutil -lc
IPO Error: unresolved : PL_comppad
        Referenced in /tmp/ipo_iccuB3iq9.o
IPO Error: unresolved : PL_regfree
        Referenced in /tmp/ipo_iccuB3iq9.o
IPO Error: unresolved : PL_regcompp
        Referenced in /tmp/ipo_iccuB3iq9.o
IPO Error: unresolved : PL_peepp
        Referenced in /tmp/ipo_iccuB3iq9.o
IPO Error: unresolved : PL_sortcop
        Referenced in /tmp/ipo_iccuB3iq9.o
...

ultimatively failing the whole compile process - of course.

[perrin]

Here is a message I sent to the mod_perl list with my results from the test.

[perrin]

Sorry, I didn't have any trouble compiling with it so I can't tell you what the issue is. I was using a 32-bit processor -- Pentium 4, I think.

[vkon]

In addition to already mentioned turning off threads/debugging, there could be a reason to try building with perl own memory allocator.

This immediately brings binary incompatibility, but you can gain dramatic speed improvements on some OSes (PocketPC is one, and win32 is also; I have not tried comparing this elsewhere)

BR
vkon