Closed pqnelson closed 1 year ago
That's a fair question, I haven't tried to measure that. Based on my experiences, the short articles are difficult to measure but rust will likely win out just because there is less file/process overhead (the rust verifier can chew through the first 20 articles in the time it takes for the vanilla verifier to do its usual progress reporting stuff for one), and there seem to be some patterns that take a lot longer in vanilla. When 5x is the baseline it seems difficult to imagine that there is an example where that goes to 1x.
Re: parser, I'm kind of hoping that this verifier gets some traction as an open source project within the Mizar community so that it's not just me doing the work. It's also the least important component re: proof export compared to accom + analyzer + checker, so it will likely be left for last if I'm the one doing it.
Okay, I had some time to sit down and do some per-file measurements for mizar-rs
vs verifier
.
verifier
:mizar-rs
, 637.738s for verifier
, 110.9x speedupmizar-rs
:mizar-rs
, 451.123s for verifier
, 3.65x speedupverifier
/ mizar-rs
:mizar-rs
, 1.499s for verifier
, 749.5x speedupmizar-rs
, 0.574s for verifier
, 114.8x speedupmizar-rs
, 637.738s for verifier
, 110.9x speedupverifier
/ mizar-rs
:mizar-rs
, 1.212s for verifier
, 0.621x (slowdown)mizar-rs
, 1.103s for verifier
0.782x (slowdown)mizar-rs
, 1.558s for verifier
1.162x speedupBCIIDEAL and LATSUM_1 are the only two data points below the $y=x$ line, which you can see on the graph.
Ignoring SUBSET and TARSKI which have absurd speedup numbers because mizar-rs
has lower fixed costs than verifier
, LEIBNIZ1 stands out as hitting a particularly bad case in verifier
resulting in a 10 minute runtime just for that one file. I have not investigated this in detail.
I will link to this post from the README for people interested in a more detailed performance comparison.
Hello,
This isn't an "issue", per se, but I was wondering if there were any files where the "vanilla verifier" from Mizar outperforms
mizar-rs
? Or is the rust version better across the board?Best, Alex
P.S. Will you continue and implement the parser, or is that "an exercise for the reader"?