[Openmcl-devel] Profiling for CCL?

[Gary Byers]

Apple stopped supporting their (generally very good) CHUD tools in OSX 
10.7.  I've heard that
people have used the general idea described in the CHUD/Shark section of 
chapter 12 to
generate a heap image with Mach-O symbol information and gotten useful 
information
out of the current profiling tools in XCode, but have never done so.  
(One of many reasons
that I still use 10.6 for most things is that I continue to find CHUD 
useful.)

oprofile is still being developed and maintained on Linux; for the last 
few years, another
Linux profiling tool ("perf") has also been available.  On 
Debian/Ubuntu, perf is available
as part of a "linux-tools" package.  Using perf with CCL generally involves:

  - loading the code that you want to profile
  - saving an image with ELF symbols prepended, as described in the 
oprofile section of chapter 12
  - using "perf record" to generate profiling information:

$ perf record ccl -I /path/to/image-with-elf-symbols -e '(progn (dotimes 
(i 1000) (fact 1000)) (quit))'

  - using "perf report" to analyze that info:

$ perf report

Both (what's left of) the CHUD support and the oprofile/perf stuff 
depend on OS support for the
on-chip performance counters offered by most modern CPUs.  Their output 
tends to be very
detailed and very low-level; they give a bottom-up view of what's 
actually going on, but many
people have said that they find it difficult to see the correlation 
between that info and code that
they're familar with.

I don't know of any similar OS support on other platforms; I once tried 
to install something that
claimed to provide such support on Windows, but got the Windows 
equivalent of a kernel panic
before I got very far in trying to use it.


Hans Huebner's advice-based profiler (available in 
"ccl:contrib;huebner;advice-profiler;") tries to
take a different approach: it wraps sampling code around existing 
definitions.   There are inherent
problems with some aspects of this (e.g., it may replace tail-recursive 
functions with non-tail-recursive
versions), but when it can be used its output may be easier for most 
people to understand.  I don't
know if that code is in fact portable to (e.g.) Windows, but don't know 
of a deep reason why it couldn't
be.  (IIRC, the code tries to use the highest-resolution timing 
facilities that it can find; I'm not sure
how hard it tries to find such a thing on Windows.)




On 12/1/13 6:26 PM, Mark H. David wrote:
> Does anyone do profiling on CCL?  What's good to use?  Saw online (http://ccl.clozure.com/manual/chapter12.html) something for Linux and Mac, two approaches.  Anything cross platform? Anything for Windows?
>
> Thanks,
>
> -Mark
>
> _______________________________________________
> Openmcl-devel mailing list
> Openmcl-devel at clozure.com
> http://clozure.com/mailman/listinfo/openmcl-devel
>