From: ser@... (Sean Russell)
Date: 2002-05-04T23:31:19+09:00
Subject: Ruby vs. Java vs. Native trivia

Hi,

I'm reposting this, because I don't think it got propegated properly the 
first time.  I can't find the post on Google, anyway.


I thought I'd post these results to the list, since you might find them 
interesting.  This really is unscientific trivia, so digest it as such.

I've been playing with GCJ, the Java-to-native compiler.  I've got this 
little application called EDT that does some basic math and generates a 
decimal-based time from the current time.  I've written it in both Ruby and 
Java, and have been using it to get familiar with GCJ.  I did some timing 
tests, just to see how things behaved, and the results are more or less 
surprising.

The important thing to note is that this is a VERY small app, so startup 
delays are disproportionately weighted.  However, this doesn't invalidate 
the tests; I have a number of little apps like this that I run multiple 
times daily, and the startup times do indeed impact the "user enjoyment" 
factor.

These are from an average of five runs, using Linux's 'time'.  edt_java is 
pure java, using the 1.4 vm; edt is the ruby version, and EDT is the 
java-compiled-to-native version:

TIMES (USER)
  ./edt_java: 1.27
  ./EDT: 0.14
  edt: 0.11
CPUS (%)
  ./edt_java: 64.33
  ./EDT: 56.33
  edt: 74.67
TOTALS (TIME)
  ./edt_java: 2.22
  ./EDT: 0.33
  edt: 0.18

TIMES: The pure java JIT doesn't get a chance to kick in on this app because 
it is so small; this is why the values reflect the pure Java version being 
ten times as slow as the native. 

TOTALS: It also comes as no surprise the pure java takes nearly 7 times as 
long as the native compiled version; that's the Java VM penalty for you.

CPU: I'm also not surprised that the native version uses the least amount of 
CPU; there's no interpretation going on.

The big surprise here is that Ruby is so fast.  It is not only interpreted, 
but is text-based, not bytecode.  Despite this, it is almost as fast as the 
native code.  It uses the most CPU -- that's the interpretation -- and if 
you normalize the CPU usage, the Ruby version is just a little slower than 
the native version, having a run TIME of 0.18s and a TOTAL time of 0.44.  
The app feels snappier because of the additional CPU usage -- I'm not 
sure why Linux decides that the Ruby version deserves more processing 
power.

Still, this is pretty amazing.  I expected the Java "native" version to do 
much better than the Ruby version.  I suspect that this is because there's 
significant overhead in the native version that wouldn't be there if I had 
hand-crafted the native version from C or ASM.

One last point: the sizes are significantly different as well.  The Ruby 
version is 2.6Kb large; the native Java version is 126Kb large, AND is 
dynamically linked to a number of libraries.  I haven't calculated how 
large the minimum run environment is (Ruby interpreter+code vs. GCJ 
libs+code).

--- SER