lth@idiotix.cs.uoregon.edu (Lars Thomas Hansen) (04/23/91)
Recently, I downloaded the Sparc binaries for MIT Scheme 7.1 from altdorf. Upon testing it, I found that performance was downright pitiful: given the usual fib program (define (fib n) (if (<= n 2) 1 (+ (fib (- n 1)) (fib (- n 2))))) and told to compute (fib 20), MIT Scheme would take over 40 seconds! Several other Scheme systems (Chez 3.0, T 3.1, Xscheme 0.16, Scheme48) took from 1 to 4 seconds on same. And no, this wasn't an isolated example of temporarily heavy load; it's repeatable. Is there something (obvious?) that I forgot to do, or simply is MIT Scheme this bad? --lars
jinx@zurich.ai.mit.edu (Guillermo J. Rozas) (04/23/91)
In article <LTH.91Apr22235109@idiotix.cs.uoregon.edu> lth@idiotix.cs.uoregon.edu (Lars Thomas Hansen) writes:
Path: ai-lab!mintaka!ogicse!cs.uoregon.edu!skinner!lth
From: lth@idiotix.cs.uoregon.edu (Lars Thomas Hansen)
Newsgroups: comp.lang.scheme
Summary: is it really that shitty, or is something weird happening?
Date: 23 Apr 91 07:51:09 GMT
Sender: usenet@cs.uoregon.edu (Netnews Owner)
Distribution: usa
Organization: /home/systems/lth/.organization
Lines: 21
Recently, I downloaded the Sparc binaries for MIT Scheme 7.1 from altdorf.
Upon testing it, I found that performance was downright pitiful: given
the usual fib program
(define (fib n)
(if (<= n 2)
1
(+ (fib (- n 1)) (fib (- n 2)))))
and told to compute (fib 20), MIT Scheme would take over 40 seconds!
Several other Scheme systems (Chez 3.0, T 3.1, Xscheme 0.16, Scheme48)
took from 1 to 4 seconds on same.
And no, this wasn't an isolated example of temporarily heavy load;
it's repeatable.
Is there something (obvious?) that I forgot to do, or simply is MIT Scheme
this bad?
--lars
The short answer is:
On a Sparc, or any other machine to which the native-code compiler has
NOT been ported, it is that slow, and I would not use it except as a
toy.
The long answer is:
The time is not spent in fib itself, but in generic arithmetic.
Generic arithmetic in MIT Scheme is huge (including ratnums, recnums,
etc.) and coded in Scheme itself (rather than in C, etc.), thus the
lack of a compiler is most noticeable in the time spent inside the
system (arithmetic, list utilities, etc.).
Generic arithmetic used to be written in C in MIT Scheme, but that was
abandoned when ratnum and recnum support was added. In the presence
of a native code compiler, that was a reasonable choice. In the
absence of it, it isn't. We could provide two different versions of
the system, one that assumes there is a compiler, and one that assumes
there is not, but it is too much work, and after all, we only make MIT
Scheme available because people want to use it, not because we want
them to use it.