|
|
|
|
|
|
by pag
3628 days ago
|
|
|
We ran three versions of our system, because in the last few weeks/days before the quals we had a bunch of regressions. One of those regressions was caused by a single line change in my fuzzer :-p On the twitter account we used to track our progress, we used rapper names for each version. One version was two weeks old, one was up-to-date with the fixed fuzzer, and one had an experimental version of one of our symbolic executors that produced <var> tags. That didn't work. We also under-utilized those nodes :-( Each node had at at least 4 idle cores wasting our money. Our resource allocation mechanism was naive. I looked through some of the stuff released byt DARPA after the event and they released some of our PoVs as official PoVs. If you hex-decode them, you'll see something like "bad seed to Radamsa"!! That was a bug in how I would invoke Radamsa -- sometimes I'd pass it a seed that was way too big. We tried to team up with every team but ForAllSecure. No one wanted to have our name on their ticket, or they were just fishing for details :-/ We've done a bit of work on the system since, getting it to work on Linux programs via a "port" of parts of libc to DECREE. |
|
|
Resource utilization is definitely tricky. But man, the amount of resources you had is just mind-boggling! I just realized it's even more cores than DARPA gave us for the final event! I'm impressed you guys managed to keep it all running smoothly (at least, it seemed that way from our lab, where there was complete chaos as our system fell over and crapped itself repeatedly for the first few hours).