Version française
Home     About     Download     Resources     Contact us    

This site is updated infrequently. For up-to-date information, please visit the new OCaml website at

Browse thread
Benchmarking different dispatch types
[ Home ] [ Index: by date | by threads ]
[ Search: ]

[ Message by date: previous | next ] [ Message in thread: previous | next ] [ Thread: previous | next ]
Date: 2007-01-19 (00:50)
From: Jacques Garrigue <garrigue@m...>
Subject: Re: [Caml-list] Benchmarking different dispatch types
From: "Nathaniel Gray" <>
> As somebody trying to understand the performance of OCaml, I've often
> wondered about the performance of the different forms of function
> dispatch.  How do method calls compare to function calls?  How about
> closure calls?  So I tried using the Benchmark library[1] to do a
> quick test:
> ========
> (* Test method dispatch vs. function dispatch vs. closure dispatch
>    Make sure to compile with -inline 0
>  *)
> let f x =
>    x + 100
> let call_f () = f 1
> let o = object
>    method f_o x = x + 100
> end
> let call_o () = o#f_o 1
> let f_c () x = x + 100
> let f_c' = f_c ()
> let call_fc () = f_c' 1
> let o_c = object
>    method f_oc () x = x + 100
> end
> let f_oc' = o_c#f_oc ()
> let call_foc () = f_oc' 1
>                     Rate       method obj. closure      closure     function
>       method  25974026/s           --          -5%         -16%         -90%
> obj. closure  27210884/s           5%           --         -12%         -89%
>      closure  31007752/s          19%          14%           --         -88%
>     function 254777070/s         881%         836%         722%           --

There are a few problems in your methodology.
One is that you are running your test only once inside a function.
So what you are measuring ends up being (at least) the cost a calling
a closure + the real cost of your test. Usually the wrapping function
should itself be a loop.
  let call_f () = for i = 1 to 1000 do ignore (f 1 + 1) done

Another problem is that with such micro-benchmarks, all kinds of
optimizations may skew results, either by the compiler or the CPU.
You disabled one with -inline 0, but there is noway to discard others
if you don't know what triggers them.

For instance, when calling a method, normally you would have to search for
it in the method list stored inside the object. This is done by a
binary search, with logarithmic cost in the number of methods in the
list. Since having to do it for every method call would badly impact
performance, each call point caches the offset in the list for the
last object called. If the last object was from the same class, then
no search is done. There are only a few extra memory reads, to verify
that indeed this is the right offset.
So if want to measure the cost in the worst situation, you have to
alternate calls (at the same point) between objects from different
classes, for which the offset is different.
In practice, hopefully this worst pattern doesn't occur too often, so
it is still safe to assume that method calls 

You should also look at the generated assembler (obtained with -S) to
verify that no strange optimization happens.

My own measurements on a Pentium M and PPC (using a slightly different
benchmark, using loops and several different methods and functions)
give (comparing to a direct function call):
                    Pentium M   PPC G4
Closure:            1.2x        3.2x
Method:             2.9x        5.6x
Unoptimized method: 6.9x        13x

I'm a bit surprised by the low cost of a closure, particularly on
pentium M, but this may be related to some CPU optimization.
Note that with inlining you get a more than 10x speedup.
This suggests that even in the best case method calls are actually
about twice as expensive as closure calls, and 5 times in a
particularly bad case.

Jacques Garrigue