[META] Stress-tests created using LLMs (Chatgpt, Grok and others) where Firefox is slow or comparatively slower.
Categories
(Core :: Performance: General, task)
Tracking
()
People
(Reporter: mayankleoboy1, Unassigned)
References
(Depends on 217 open bugs, Blocks 1 open bug)
Details
(Keywords: ai-involved, meta)
This is a meta-bug to track all the stress-tests I create using LLMs like ChatGPT. The testcases span across different components, so this META in Core:General seems appropriate.
Updated•1 year ago
|
Updated•1 year ago
|
| Reporter | ||
Updated•11 months ago
|
Comment 1•10 months ago
|
||
I wonder we could use our fuzzer infra to generate this kind of perf tests, because it looks like we are basically doing the same - generating random scripts, run them, and report in case there's an issue. Generating tests from our own Web IDL would be able to cover new APIs that LLM is not yet trained for, or to not cover what Firefox does not support.
Comment 2•10 months ago
|
||
This bug has over 100 dependencies, and most of them are still open, so I'm not sure there's a big rush to generate more bugs that we're not going to fix. Finding quadratic behavior in the browser is fun, but I'm not sure whether we should prioritize it if no actual web pages are hitting it.
| Reporter | ||
Comment 3•10 months ago
|
||
I can share what i have seen since the time i started this:
- The "number of iterations" is based on trial and error. For some APIs/functions/calls/whatever, number as low as 5000 is enough to show bad behaviour. For others numbers as large as 50000000 is needed. It depends on how the browser behaves for that APIs/functions/calls/whatever.
- In many cases, the browser will freeze for multiple minutes or use uncontrolled explosive memory. Or it will do neither, but will just use 100% of one CPU for multiple minutes which doesnt freeze the browser nor does it uses memory. The only way i have seen is to check the resource usage on Task Manager.
- What is the definition of 'bad' ? Whether bad in absolute terms, bad in comparison with other browser(s) ? Bad for realistic iteration counts? What is realistic iteration count?
- Some of the bugs which are "bad" are uninteresting because the iteration count is pathological, completely unrealistic. What would be the filter criteria for these bugs? If every issue that the fuzzers find are filed as bugs, i suspect the number of bugs will become explosively high.
- Root of a lot of bugs i have found are around (AFAIU) cleanup/cc-type-things, backpressure on IPC/IPDL, nsTarray ... These are sort of known in the sense that these are tuned towards realistic browser use (or realistic browser benchamrks like SP3). Optimizing for these extreme cases may be bad for the general use, if possible at all.
- IF we go the route of fuzzers doing this, these known issues must be fixed first. Else i suspect majority of bugs would be around these things only.
Comment 4•10 months ago
|
||
I think often times benchmarks log operations-per-second rather than the total time, rather than setting an iteration number and run the full iterations. That would solve many listed issues, except 3 - that still needs some definition.
Comment 5•10 months ago
|
||
(In reply to Kagami Rosylight [:saschanaz] (they/them) from comment #1)
I wonder we could use our fuzzer infra to generate this kind of perf tests, because it looks like we are basically doing the same - generating random scripts, run them, and report in case there's an issue. Generating tests from our own Web IDL would be able to cover new APIs that LLM is not yet trained for, or to not cover what Firefox does not support.
This is something the fuzzers can do in some capacity and we have filed bugs like this in the past. The issue we've had is that there's varying level of appetite for fixing bugs like this.
| Reporter | ||
Updated•9 months ago
|
Updated•9 months ago
|
Updated•8 months ago
|
| Reporter | ||
Updated•3 months ago
|
| Reporter | ||
Updated•1 month ago
|
| Reporter | ||
Updated•1 month ago
|
| Reporter | ||
Updated•1 month ago
|
Description
•