Skip to content

PERFORMANCE.md

Latest commit

 

History

History
120 lines (101 loc) · 2.44 KB

PERFORMANCE.md

File metadata and controls

120 lines (101 loc) · 2.44 KB

Benchmarks

We haven't yet focused on trying to make Latexjs fast. Here's some initial benchmarks:

time node ./pdflatex.js -synctex=1 -interaction=nonstopmode -output-format pdf /app/demo.tex

real    0m2.064s
user    0m1.964s
sys     0m0.120s

If built with --memory-init-file 1:

real    0m2.005s
user    0m1.896s
sys     0m0.128s

time ./texlive/bin/x86_64-linux/pdflatex -synctex=1 -interaction=nonstopmode -output-format pdf /app/demo.tex

real    0m0.260s
user    0m0.220s
sys     0m0.040s

8x. ouch.

With a larger file (a 20 page paper of mine)

time node ./pdflatex.js -synctex=1 -interaction=nonstopmode -output-format pdf main.tex

real    0m4.213s
user    0m4.072s
sys     0m0.176s

time ./texlive/bin/x86_64-linux/pdflatex -synctex=1 -interaction=nonstopmode -output-format pdf main.tex

real    0m0.583s
user    0m0.528s
sys     0m0.052s

So we are about a factor of 7x slower...

And on a much larger document (my thesis):

real	0m6.575s
user	0m6.220s
sys	    0m0.454s

real	0m2.640s
user	0m1.307s
sys	    0m0.360s

which is 2.5x slower.

This suggests that actually the warm up of the JIT is a significant factor - the longer we run for, we amortise this cost and we approach a 2x perf penalty.

Preliminary WASM results

See the WASM branch for how I build a version using WASM. Have to use bleeding edge node build, see latexjs#1

Firstly a quick cf with pdftex on Linux

time ~/Dropbox/latexjs/bin/node pdflatex.js -synctex=1 -interaction=nonstopmode -output-format pdf main.tex

real    0m1.870s
user    0m3.300s
sys     0m0.156s

time pdflatex -synctex=1 -interaction=nonstopmode -output-format pdf main.tex

real    0m0.525s
user    0m0.476s
sys     0m0.044s

3.56x, which is around 2x faster than non-WASM (see above)!

However, I notice the start up is quite slow:

time ~/Dropbox/latexjs/bin/node pdflatex.js --version

real    0m0.907s
user    0m2.328s
sys     0m0.136s

time pdflatex --version

real    0m0.005s
user    0m0.004s
sys     0m0.000s

Ouch, a full second startup penalty..,

thinfs_db.json is pretty big, how expensive is loading that? I cut it down to a minimal version (one record)

real    0m0.579s
user    0m2.004s
sys     0m0.104s

Much better! This implies keeping a small cache thinfs_db file could save 0.3s..., which would bring us closer to 3x perf delta.