Hacker News Reader

Comment by bjourne | Hacker News Reader

I'm referring to Takikawa et al (2016), who reported performance 50-100 times worse performance numbers by mixing typed and untyped Racket code together. Actually 100 times slower execution. They used type annotations on the module level. Would be interesting to know if Racket developers have proven them wrong and if so, how.

http://www.ccis.northeastern.edu/home/types/publications/gra...

View on HN

Just to make it clear, the authors of that paper (including myself) are all to some extent Racket developers. Some big improvements have been made but there are still pathological cases. For the latest published on this see this paper: http://users.cs.northwestern.edu/~robby/pubs/papers/oopsla20...

But also it's important to note that it's not Typed Racket or Racket in isolation that are slow, but the inter-mingling of the two due to contract overhead.

Yes, that fits with my own experience. I've not used Typed Racket itself, but I've used contracts in untyped Racket. I found them so slow that I ended up using a macro which discarded them unless it was a run of the test suite.

Figure 3 in that paper is enlightening: the fully typed version takes 0.7x as long as the untyped version, so Typed Racket is slightly faster than normal Racket. Most of the partially-typed versions take 50x to 100x as long, as you say, showing that it is indeed the contracts that slow everything down.

Takikawa is a core racket dev, I believe.

Reply on news.ycombinator.com