Re: [PerfSuite-users] psrun fatal errors: hardware event not available

[Rick K. <rk...@nc...>]

Harry,

> Thanks for the zippy reply - much appreciated.

You're very welcome, although I'm sorry the reply didn't exactly offer a 
current fix!

> Just a suggestion - one of the tests that should run (maybe it does and I 
> missed it) is one that checks for the kernel version.  It seems odd that NONE 
> of the tests fail (I thought that originally the threads test failed, but in 
> the output below, they pass).  Seems like a pretty good test result for 
> something that won't run :)

That's a good idea and as you point out, no such check exists in the 
0.6.2a2 releases or earlier.  The upcoming version does have a check for 
kernel major and minor versions in the configure script, not only to 
possibly bail on error but also because there are a few things for which a 
different code path has to be taken based on kernel-specific things (like 
in /proc).

> I read the Linux Journal article which had no mention of the 2.6 kernel 
> prohibition and I must have missed the part where this was noted.  In the 
> INSTALL, it mentions:
> 
>     124       Currently, the latest 2.6.x release of perfctr is the best 
> choice of
>     125       versions for new installations of the driver.  Please read the
>     126       instructions in the perfctr distribution carefully.
> 
> which implies that the 2.6 kernel is supported.

Yes, this is kind of confusing.  The release numbers of perfctr, which is 
what is being referred to in the INSTALL file, are not in lockstep with 
the kernel versions.  That is, you can run perfctr 2.6.x on a 2.4.x 
kernel.  Also, you didn't miss anything in the LJ article... it just 
wasn't addressed :(

> I can't argue with free SW, and I'm trembling with antici.....pation :) for 
> the 2.6 supporting version, but it might have been made clearer.

I agree.  The web sites *and* the INSTALL instructions need a revisit.  
And, of course, a newer release for current kernels will solve the real 
issue :) 

> If there's currently a cvs/svn version of perfsuite, I'd be willing to give it 
> a try.

Thanks much for the offer.  There isn't a publicly-accessible CVS tree, 
but the next (hopefully final alpha) release is on its way.  What is going 
on right now are updates for the Perfmon (IA-64) driver (which *is* tied 
to the kernel major version).  Other generic 2.6 updates are in place but 
it makes sense to package it all together.

> Also, the difference between root and user was only due to permissions 
> on /dev/perfctr.  changing to a+r allowed a normal user to obtain the same 
> results.

Ah, that's good to know.  Thanks for clarifying.

> Do you have any idea if the new version of PAPI will allow perfsuite to run on 
> 2.6 (ie is the reason it won't run due to PAPI or due to perfsuite?)

It's not a PAPI-specific thing, it's in PerfSuite.  Specifically, 
necessary mods are in libperfsuite, not libpshwpc.  It is possible that 
psrun, in counting mode, might run OK on a 2.6 kernel with PAPI 3, but 
it's been a while since I checked.  I'm certain that profiling would be 
affected, though, and that would be the case whether profiling with 
hardware counters or with interval timers or profil().  Counting with 
Perfmon directly also would fail (but that is not applicable to Opteron at 
this point in time).

I've lopped off your test suite results (thanks for sending though).  The 
test suite is best at detecting fatal (core dump-inducing) type errors and 
has to be a little generic, because it has to be able to run whether or 
not the kernel has been patched and PAPI is installed.  But you can check 
whether you're getting anything meaningful at all through PAPI like this:

$ cd src/libpshwpc/tests
$ ./testsuite -v -k count

You should get "verbose" output from the test cases that try to count 
events, and if the results are reasonable (non-zero), you might be OK 
(again, for counting only... not profiling).

Thanks again for the feedback,
Rick