SML and OCaml: Why was OCaml faster?

http://thebreakfastpost.com/2015/05/10/sml-and-ocaml-so-why-was-the-ocaml-faster/

by cannam • 11 years ago

83 22 11 years ago

thebreakfastpost.com

saosebastiao• 11 years ago

So basically Ocaml was faster because it was FFI?

ori_b2• 11 years ago

Both did FFI. Ocaml was faster because there was less data munging around the FFI.

agumonkey2• 11 years ago

Still surprising how much time is devoted to formatting (ocaml or any language).

ori_b2• 11 years ago

Input and output of floating point involves a decent amount of bigint math to do correctly. It's not surprising to me that it is slow -- doing it quickly is still an active area of research.

Writing a bit about that is still on my todo list.

agumonkey2• 11 years ago

Agreed, but it's mostly unnecessary on the logic side, maybe communicating through binary forms to avoid transcoding costs.

thechao• 11 years ago

I remember reading, years ago, a paper which provided an algorithm for converting an IEEE 754 into a string using only a fixed number machine word length scalars—but I certainly can't find it now.

CHY8724• 11 years ago

This is why we don't microbenchmark languages. Usually if you dig a little deeper, you find that actually the difference between the languages is mostly just that one of them does one thing badly.

This by the way is MLton's worst case scenario. MLton is designed to perform full-program optimisation (unlike almost any other compiler I've seen); the gap is supposed to grow as the code gets larger (read: more realistic).

Sadly due to its design tradeoffs (it's a compiler) it's totally unsuitable for the big ML projects (interactive theorem provers).

jasperry2• 11 years ago

Can you elaborate on why being a compiler makes MLton unsuitable for theorem provers? Is it because they are usually written on top of the interpreter's REPL, and need its reflection capabilities?

loup-vaillant• 11 years ago

I'm guessing it's the significant compilation times required by whole program optimisations. MLton doesn't do separate compilation.

(I think. Citation needed. It may separate some things, then optimise after, GCC -flto style.)

CHY8724• 11 years ago

It's because they use metaprogramming so extensively. As an example, HOL4 (I can only speak for that) accepts a script written by the user (as though in a REPL), compiles and runs it, and produces an sml theory file which represents their defined theory [0]. The user when they start their next theory are supposed to use this computer generated code.

You would have to write your own repl, write your own way of loading, writing theories etc and by that time you've lost performance (HOL even after 20 years of progress and tweaks now has only mostly ok performance).

So it would at least require some rearchitecting, which is my point; it's not a drop in replacement.

[0] - This is AES, before and after building! https://gist.github.com/j-baker/244ef3e59f19d74c352b

more_original2• 11 years ago

MLton is unsuitable for LCF-style theorem provers, simply because it has no interactive toplevel.

ML was originally developed as a MetaLanguage for such theorem provers. The idea is to have an abstract type thm of theorems, whose operations (like modus ponens) only allow you to construct provable theorems. To prove a theorem interactively, you would start the ML toplevel and try to construct a proof of your proposition by programming in ML. The ML type system guarantees that you can only prove valid things. The toplevel is essential if you want to prove things interactively.

Modern theorem provers like Coq work differently. If Coq were written in SML rather than OCaml, then MLton should be good for Coq as well. I think people have experimented with non-interactive versions of HOL4, compiled using MLton (because it's faster than other SML compilers).

dman• 11 years ago

You might want to check out Stalin. (http://en.wikipedia.org/wiki/Stalin_%28Scheme_implementation...)

tsuyoshi• 11 years ago

Another issue, comparing the MLton FFI and the Ocaml FFI, is that the FFI in MLton is somewhat slower. When I converted an Ocaml program to SML, I discovered that a small portion written in C ran much faster when I rewrote it in SML. In Ocaml, writing portions in C is almost always a performance win, just because the Ocaml compiler doesn't optimize very well. In MLton there's a bit of overhead just for the FFI (possibly because the code generator allocates registers very differently from how it's done in C). That overhead goes away if you use the C code generator instead, but usually the native code generator performs better anyway.

If I remember correctly, it has been proposed to rewrite the MLton floating point conversion in pure SML for precisely this reason.

SML and OCaml: Why was OCaml faster?

Related Stories