[BUG] js_of_ocaml is excessively memory hungry

[JasonGross]

Describe the bug

js_of_ocaml is very cool! I use it on CI to generate a webpage. However, I cannot use it on the new GitHub Actions arm64 MacOS boxes, which have only 7 GB of RAM, because it sometimes eats 8--9 GB RAM to generate a single .js file. For example here is a table of build times and memory usages on linux:

     Time |   Peak Mem | File Name                                         
---------------------------------------------------------------------------
15m11.31s | 8904232 ko | Total Time / Peak Mem                             
---------------------------------------------------------------------------
 4m46.85s | 8808932 ko | ExtractionJsOfOCaml/bedrock2_fiat_crypto.js       
 4m37.20s | 7151532 ko | ExtractionJsOfOCaml/fiat_crypto.js                
 4m30.42s | 8904232 ko | ExtractionJsOfOCaml/with_bedrock2_fiat_crypto.js  
 0m25.80s | 3195952 ko | ExtractionJsOfOCaml/with_bedrock2_fiat_crypto.byte
 0m25.32s | 3193600 ko | ExtractionJsOfOCaml/bedrock2_fiat_crypto.byte     
 0m24.60s | 2814364 ko | ExtractionJsOfOCaml/fiat_crypto.byte              
 0m00.38s |  102780 ko | ExtractionJsOfOCaml/bedrock2_fiat_crypto.cmi      
 0m00.37s |   98468 ko | ExtractionJsOfOCaml/fiat_crypto.cmi               
 0m00.37s |  102360 ko | ExtractionJsOfOCaml/with_bedrock2_fiat_crypto.cmi 

It is similar on mac, and a bit better on debian sid.

I invoke it with --source-map --no-inline --enable=effects and invoke the compiler with -package js_of_ocaml -package unix -w -20 -g

For the near future (until artifacts expire), the build artifacts page contains generated .js files (fiat-html-js-of-ocaml), .ml source files (ExtractionJsOfOCaml-source-master), and compiled files (ExtractionJsOfOCaml-master-ocaml-4.11.1).

Expected behavior
I expect there to be a way to make the js_of_ocaml pipeline fit in under 7GB of RAM, possibly with a flag, if necessary.

Versions
js_of_ocaml 5.7.2, ocaml 4.11.1

[hhugo]

• --no-inline is no longer necessary (since js_of_ocaml.5.7.0).
• dealing with debug info seems to be slow in your use case. Removing --source-map should speed up your build. I'll try to investigate.
• adding --disable globaldeadcode should give you a good speedup as well.

With the change mentioned, I the following for ExtractionJsOfOCaml/bedrock2_fiat_crypto.js

112.97user 1.26system 1:54.24elapsed 99%CPU (0avgtext+0avgdata 4581104maxresident)k
0inputs+21904outputs (0major+1275732minor)pagefaults 0swaps

[hhugo]

with #1614, one no longer need to disable globaldeadcode, at the cost of extra memory usage.

78.16user 1.85system 1:20.13elapsed 99%CPU (0avgtext+0avgdata 6034576maxresident)k
0inputs+19696outputs (0major+1631738minor)pagefaults 0swaps

I'll try to improve #1614

[hhugo]

I've updated #1614, I now get

73.86user 1.00system 1:14.87elapsed 99%CPU (0avgtext+0avgdata 4581396maxresident)k
0inputs+19696outputs (0major+1209270minor)pagefaults 0swaps

I still need to investigate the sourcemap issue.

Can you test #1614 and confirm it solves part of your issue ?

[OlivierNicole]

@hhugo Regarding this, I have been working on the sourcemap slowness and will open a PR today.

[hhugo]

@JasonGross, any luck with #1614 ?

[JasonGross]

I have not had a chance to try it, but if it works on your end, I don't see why it would be any different on GitHub Actions. (The files I linked to are the ones I actually use, not simplified examples .). But I can set up GHA to use the PR. Should I just clone the repo and opam pin add . on that branch?

[OlivierNicole]

#1617 may help too if the memory consumption happens to be in sourcemaps. I find that it nearly halves the peak memory usage when linking JSOO using itself.

[hhugo]

@JasonGross, I think you can just do

opam pin add js_of_ocaml-compiler https://github.com/ocsigen/js_of_ocaml.git#speedup

[JasonGross]

I've set it up on CI:
pre (specifically here)
post speedup (mit-plv/fiat-crypto#1922) (#1614) (still in progress)
post optim_sourcemap_link (mit-plv/fiat-crypto#1923) (#1617) (still in progress)

[OlivierNicole]

It looks both #1614 and #1617 make both the run time and the peak memory usage worse. I haven’t worked on #1614 but that surprises me a lot in the case #1617. Are these tests runnable on Linux? I may try to inspect them locally.

[hhugo]

Not all CI jobs have been updated.

Here is what I see for #1614

1m28.92s | 4461616 ko | ExtractionJsOfOCaml/with_bedrock2_fiat_crypto.js

[hhugo]

And for #1617

4m47.42s | 5253024 ko | ExtractionJsOfOCaml/with_bedrock2_fiat_crypto.js

[hhugo]

compared to

6m08.88s | 7720996 ko | ExtractionJsOfOCaml/with_bedrock2_fiat_crypto.js

[hhugo]

@OlivierNicole, I would expect your PR to only affect separate compilation during the link step but I don't think separate compilation is involved here. What part of your PR would improve the situation during whole program compilation ?

[OlivierNicole]

Not all CI jobs have been updated.

Here is what I see for #1614

1m28.92s | 4461616 ko | ExtractionJsOfOCaml/with_bedrock2_fiat_crypto.js

I didn’t quite follow which of the many jobs to inspect to find the info, but I trust that your numbers are right.

@OlivierNicole, I would expect your PR to only affect separate compilation during the link step but I don't think separate compilation is involved here. What part of your PR would improve the situation during whole program compilation ?

I’m honestly not sure. Looking into it.

[OlivierNicole]

I switched to `Stringlit  and Yojson.Raw (rather than `String and Yojson.Basic) because it saves a non-negligible amount of time on the parsing and the writing of the mappings fields of source maps (essentially, Yojson.Basic.to_string (`String s) checks for special characters or Unicode code points in the string, which takes a suprising amount of time and is unnecessary on mappings since they contain only base64 numbers, commas and semicolons.

I’m not sure it explains it all though. Trying to profile locally.

[JasonGross]

Are these tests runnable on Linux?

Yes. The cheapest way to run them is to download any of the artists labeled ExtractionJsOfOCaml-source* from our CI. These artifacts contain a handful of self-contained .ml files, the ones that we want to turn into .js files. I gave the flags I use in the initial post.

The expensive way to run the tests is to clone the repo, do opam install coq, and then run something like make js-of-ocaml

[OlivierNicole]

I can’t reproduce a significant difference in terms of run time nor profile between master and #1617. The time spent on source maps is negligible compared to the time spent optimizing. I’m starting to suspect that the CI run times have a huge variance.

[OlivierNicole]

P.S. I’ve done the test on with_bedrock2_fiat_crypto. Peak memory usage is not significantly affected, either.

[hhugo]

#1614 has been merged. Reopen if you still have issues