"make test" WARNING

Open discussion of PAPI.

"make test" WARNING

Postby KotasCW » Mon Aug 27, 2012 4:11 pm

I am in the process of installing and verifying papi 5.0.0 on an Intel Xeon E5530 computer running Ubuntu 11.04 (uname -r gives 2.6.38-11-generic). I am new to using papi hardware counters and have only a vague idea about how they work. Please bear with me.

Everything appears to build without issues. When I follow it with:
> make test
To run the ./ctests/zero routine, the value given for PAPI_FP_INS is 59999989. This gives a warning message (Warning: FLOPS validation). The expected value is 40000000. (For completion, the other counter, PAPI_TOT_CYC, is 200020342 which is ballpark close to the reported Real cycles of 180924135).

The test is listed as having "PASSED with WARNING".

Questions:
1. Is a 50% overage acceptable/expected?
2. Has something gone wrong in my install?
2a. If so, what additional information do I need to debug this install?
3. Could papi be counting something running in the background as well as the "zero" test process (which would artificially inflate the PAPI_FP_INS)? (For the record, I don't think there are any other active processes, but I wonder about how careful I need to be in the future.)

I appreciate your help. Thanks!

Additional information:
1. My compilation procedure was:
> cd <papi-root>/src/components/cuda
> ./configure --with-cuda_incdir=/usr/local/cuda/incude --with-cupti_incdir=/usr/local/cuda/extras/CUPTI/include --with-cupti_libdir=/usr/local/cuda/extras/CUPTI/lib64
> cd ../../
> ./configure --with-components=cuda
> make

2. This does not appear to be related to the CUDA component since compiling it without it produces a similar output. I do not think this is limited to PAPI 5.0.0 since the 4.4.0 library also produces a similar output (but does not give a warning message).

3. Running "make fulltest" gives mostly "PASSED" outputs, with the following exceptions:
ctests/all_native_events:Segmentation fault
ctests/cycle_ratio:cycle_ratio.c gives 2 warnings (Line # 92, PAPI_REF_CYC = 0 and Line # 167, validation)
Between ctests/krentel_pthtreads and kufrin.c there is a Segmentation fault message
ctests/profile.c gives a warning at Line # 143 (PAPI_PROFIL_RANDOM not supported)
KotasCW
 
Posts: 3
Joined: Mon Aug 27, 2012 3:25 pm

Re: "make test" WARNING

Postby danterpstra » Mon Aug 27, 2012 4:35 pm

This looks about right. Wikipedia tells me that E5530 is a Nehalem. I would have expected floating point to be better for that architecture than what you're seeing, but I know that on Sandybridge it's a major crapshoot. Given the almost exact 50% overshoot, I would suspect that what you're seeing is an artifact related to inappropriate counting of floating point load/store operations. The good news is that this number should still be roughly proportional to the "real" answer; the bad news is that there's probably little we can do to make it better. If you're truly curious, you could try devising a test to chain arithmetic operations together to increase FP operations without increasing loads and stores. See what happens to the count vs. theoretical.
One of the issues we're dealing with is that in PAPI 5.0 the error testing on zero.c is now more rigorous, thanks to input from Phil Roth. In 4.x you wouldn't have seen this warning at all. Thank him for us next time you see him :)
As far as the other tests are concerned, segmentation faults are never a good thing, but are probably PAPI issues, not problems with your installation. It would probably be worthwhile investigating those in a little more detail, particularly the all_native_events test.
cycle_ratio only works on newer kernels (>3.3, I think), so warnings there are expected.
The lack of support for PAPI_PROFIL_RANDOM is also expected.
I hope this helps.
danterpstra
 
Posts: 63
Joined: Wed Jun 23, 2010 2:21 pm

Re: "make test" WARNING

Postby KotasCW » Tue Aug 28, 2012 7:35 am

Thank you for the feedback! It sounds like I am good to go then. Now for the fun part where I try things and see if they work.

I'll try to start another thread for the all_native_events segmentation fault, since it is likely a separate issue.
KotasCW
 
Posts: 3
Joined: Mon Aug 27, 2012 3:25 pm


Return to General discussion

Who is online

Users browsing this forum: No registered users and 0 guests

cron