fuzzbench
fuzzbench copied to clipboard
Make FuzzBench more robust wrt completing experiments
Currently it's possible for experiments to never finish.
@jonathanmetzman can you provide more details here on what's going wrong and how we can fix it?
I can't seem to find a single recent one on https://www.fuzzbench.com/reports/experimental/index.html that seems to have completed successfully :/
@jonathanmetzman can you provide more details here on what's going wrong and how we can fix it?
FuzzBench assumes that every trial finishes. When you give it fuzzers that don't finish fuzzing (because they OOM etc) they break this assumption and measuring will never terminate. I guess we can fix it by passing each measuring thread/proc a bool on whether the trial ended without being preempted. If we can't find a corpus in this case, it's pointless to retry measuring the same trial (as we currently do).