Parallelising the analysis

[olivierpierre]

An analysis currently runs as follows:

1. Launch the program under test once and trace with strace every system call made;
2. For each system call traced, run the program twice a) stubbing and b) faking the system call in question.

Step 2 is obviously the longest and lends itself well to being parallelised.

[nimit]

Hi @olivierpierre, can I work on this?

[olivierpierre]

Hello, sure! This has the potential to speed up the analysis significantly. Please feel free to ask any questions here, I'll try to help as much as I can.

Thank you!

[nimit]

Hi @olivierpierre,
I have reviewed the source code for loupe and I have a parallelization approach in mind.

Here, we have an opportunity to use Python's multiprocessing library to spawn two processes - one running the benchmark for stubbed the system calls and another running the benchmark for faked system calls.

loupe/explore.py

Lines 875 to 879 in 17323db

	stubbing_perf = explore_perf(str(ERRNO_ENOSYS), probably_works_stubbed)
	print_perf(stubbing_perf, baseline_perf)
	info("Gathering data for faking...")
	faking_perf = explore_perf("0", probably_works_lying)

We can further parallelize inside the explore_perf function where we could have a pool of workers (number of workers in the pool would most likely be a configurable parameter, with default equal to NO_RUNS_AVG) that can execute the runs in parallel.

loupe/explore.py

Lines 456 to 465 in 17323db

	for j in range(NO_RUNS_AVG):
	log = get_temp_file()
	success = False
	while (not success):
	with open(log, 'wb') as logf:
	process = start_seccomp_run(errno, [i], logf,
	prefix=["taskset", "-c", str(TASKSET_CPU)])
	testcmd = [testscript_path, log, "benchmark"]

How does this sound?

[olivierpierre]

It looks good to me.

I think it's a good idea to go with gradually increasing level of complexity (i.e. of parallelism) so it makes sense to first try to parallelisme the exploration of faking syscalls with that of stubbing syscalls. Then later you can try parallelising the faking/stubbing of each system call.

[Th0ught09]

@nimit how do you make those code blocks, they're rather neat :)

[nimit]

@olivierpierre got it!
I will create a draft PR

[nimit]

@Th0ught09 thanks :)
Edited 5 times before the code blocks showed properly 😂
In the source file, click the line number and shift-select another line number to select all the code between those lines. Right-click to get a permalink and paste it in comments/PRs to automatically get the code blocks. You can follow this official guide if you get stuck.