Sunday, 28 April 2002
Copyright © 2002 Gene Michael Stover. All rights reserved. Permission is granted to copy, transmit, store, and/or retrieve electronically.
I was learning Perl the other day. For lack of anything else to program initially, I write a few useless tiny programs. Since they didn't do anything useful, naturally I timed their execution. What's the point in knowing how long they required to run if I didn't know how long similar programs in other languages required to run? So I wrote them in Lisp, too. Then in C. Before I knew it, I had written speed tests in just about every language I see every day. (All except Bourne shell, I guess.)
So here, for you, are the results from one of those test programs.
A test program computes a Fibonacci number many times. It uses the definitive, but naive, recursive algorithm. That's right; said algorithm is grossly inefficient, repeating calculations which don't need repeating. That makes it perfect for a CPU-bound performance test that doesn't require a lot of memory.
Some people will take joy in pointing out that a single test, especially such a simple one, cannot be representative of the performance of a whole language. Those people would be right. Those people would also be annoying, so if you are such a person, just read the numbers, think about how naive I am, & keep your opinions to yourself, or at least away from me.
I ran these tests on palsy.CyberTiggyr.COM. It's an
Intel Pentium whatever with 128 megabytes of real memory,
running OpenBSD 2.7.
C & C++ programs were compiled with Gnu's gcc
version 3.0.4. The Java compiler & run-time were from JDK 1.3.1
for Linux, which I obtained from Sun's Java site. Lisp was
clisp version 2.28. Perl was version 5.6.1. Only the
good Lord above, or the bad one below, knows that version of Python
I used.
No other users were logged in, & I wasn't running any other programs at the time.
Here are some results. (All in separate files of their own.)
| iterations | F in "(fib f)" | report |
|---|---|---|
| 1 | 20 | results-0001-20.html |
| 1 | 25 | results-0001-25.html |
| 1 | 30 | results-0001-30.html |
| 1 | 35 | results-0001-35.html |
| 1 | 37 | results-0001-37.html |
| 1 | 40 | results-0001-40.html |
You might notice that fibb, the Bourne shell
program, has a showing only in tests with low F.
That's because it's a couple orders of magnitude slower than the
others.
You can create your own results, too. The program fib-go runs all the tests & puts their results in
an HTML file which it prints to STDOUT. Just run
fib-go, redirect the output to your own file, &
load that file into your Web browser. Control the number of
iterations with a required I command line argument
& with a required F argument. For example,
"fib-go 10 20" would run 10 iterations of
fibonacci(20) for each language.
You can download the whole kit-&-kaboodle containing this article, my files of results, & the programs, from my page at Freeshell. That's speed-fib.tar.bz2.
C & C++ are effectively the same. This test didn't use any special or object-oriented features of C++, though.
Java differs from C/C++ by a surprisingly small amount. It looks like the cost of Java is about 1.4 that of C, so it's the same order of magnitude as C. That's without any object-oriented features of Java, though.
Lisp is two orders of magnitude slower than C/C++.
Python & Perl are nearly three orders of magnitude slower than C/C++.