Golden Output at profiling stage incorrect and nondeterministic

DependableSystemsLab / LLFI-GPU

Fault injector for GPUs based on the LLFI Fault Injection Tool

Other

15 stars 2 forks source link

Golden Output at profiling stage incorrect and nondeterministic #4

Open beckdav opened 7 years ago

beckdav commented 7 years ago

Hi,

I have LLFI-GPU running on Ubuntu 12.04, LLVM 3.0 and CUDA 6.0,

I have a (tried and tested) CUDA implementation of an deterministic algorithm I wish to analyse with LLFI-GPU, this algorithm outputs a binary data file which is my golden output.

However when compiled with LLFI-GPU profiling steps the golden output file of my CUDA software is both incorrect and nondeterministic (changes each time the executable is run), (profiling information output seems to be correct).

I do not understand why this behaviour is occurring, any suggestions would be greatly welcomed!

cclinus commented 7 years ago

Can you point us your benchmark using ? and its makefile too ? You need first make sure that your benchmark application is deterministic...

Thanks

ckalra commented 6 years ago

Hi, I am having the same issue on K20 GPU with the same machine configuration mentioned by beckdav.

I tried the Blackscholes benchmark from CUDA SDK. Source is attached. After placing the files in the LLFI directory, update the path of include directory in profile.py file flagHeader, before running 'python profile.py'. If you run the benchmark without the profiler, it passes the test. Please let me know if you're able to reproduce it on your machine.

Thanks. llfi_issue.zip

abanwer commented 5 years ago

Hi, I have looked into this issue. The case seems to be specific to this bench mark, I have tried this with other benchmarks and LLFI-GPU seems to be working fine. I am looking into this issue. Meanwhile you can try using LLFI-GPU with some other benchmark to check your results. I'll let you know when I have any update or fix.