ooprolog: Running out of memory

[agentsim]

After following the recommended sequence of commands for analyzing large executables (mine is 14M) I wind up with a 105M ooprog-facts.pl file.

Using the suggested command, ooprolog --facts ooprog-facts.pl --results ooprog-results.pl --log-level=6 >ooprog.log, ooprolog runs until ultimately it is killed by docker after consuming all 128G of RAM on my system.

I've tried using the switches to limit the stack and table sizes but that did not reduce total memory use. I'm running on an M3 mac, so I also rebuilt the docker image for ARM64 CPUs, that did not help.

Generally I see about 1350 guesses before being out of memory. Is there any way to know how close to complete the process is? Is this amount of memory expected for an executable of this size?

[sei-eschwartz]

Wow, that is quite large, but we have seen people run OOAnalyzer successfully on executables of that size. On the other hand, we've also seen it fail on executables of that size.

I would guess that 128G is on the border of what is necessary. I would suggest making sure that all of your RAM is actually being exposed to Docker. Normally it defaults to something much smaller (8G or something).

If you can share your executable, I can try to run it on our system and see if there are any obvious problems.

[agentsim]

I cannot share the executable, but it seems to me that ooprolog doesn't require the executable to run. I could send you the ooprog-facts.pl file to run on your system if that helps.

Yes, I have configured docker to use the full 128G of RAM in my system. I've also tried to convince it to use swap, but that hasn't worked.

[sei-eschwartz]

I can do less with the facts file, but sure, you can send that to me at eschwartz@cert.org

[sei-eschwartz]

I am running this now. When running sanity checks during the initial reasoning process, a problem is detected (#283).

With that disabled, I just reached the guessing stage. Memory use is ballooning dramatically. It was somewhere around 20 GiB before guessing. It's now at 55GiB and still hasn't even made a guess. So it's obviously tabling something that is enormous.

[sei-eschwartz]

So my run also crashed on a machine with 256 GiB of memory. I'll try to run using some SWI Prolog profiling tools, but I'm not too hopeful.

[agentsim]

What causes such a huge amount of memory to be used on an executable that is many orders of magnitude smaller?

[sei-eschwartz]

It is from tabling (basically memoization). I will know more soon. But my guess is that there is a table that is basically Function*Function, and because there are so many functions, the table size is ridiculous.

[sei-eschwartz]

I have been running the example using https://github.com/SWI-Prolog/webstat but I have been running into a problem. The tables grow so large that when I try to get information on them, the request times out. I'm trying to figure out where this timeout is specified.

[sei-eschwartz]

I think it's from https://www.swi-prolog.org/pldoc/man?section=httpserver#:~:text=time_limit(%2BSpec,setting/2.

[sei-eschwartz]

Just a brief update, I set this to 1800 seconds, and still hit the timeout. I'm probably going to have to reach out to the SWI Prolog maintainer to see if there is any way to improve the performance.

[sei-eschwartz]

I reached out to Jan here