Re: fun with Java

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA256

Matthias B. wrote:
> Was the bytecode program compiled by GCJ, too? What about if you compile
> it with Sun's javac? Any difference?

The native program was generated by compiling the java source code with 
GCJ using -O4 optimization.  The bytecode executed by Sun's VM was 
generated by compiling with Sun's javac using -O optimization.

I think at least part of the problem with hotspot may be the internal 
representation of strings.  I noticed that even eliminating the 
tokenization step and just reading the messages from the mbox file into 
memory one at a time (i.e. each message loaded into memory in its 
entirety then discarded) caused the Sun VM to use a ton of memory.

It was also interesting to watch the resident set size fluctuate in top. 
  With spamprobe and the GCJ compiled version the RSS did change much 
(maybe 10% or so) once it hit its average value.  With the Sun VM OTOH 
the RSS fluctuated wildly from 40-400 MB!  Of course top is a very 
course grained way of watching memory usage but the runs were very 
consistent.  It appears that the Boehm garbage collector used by GCJ 
puts a much smaller hit on the processor cache than the collector in 
hotspot.

Overall I think gcj is pretty great and I'd very much like to explore 
compiling some of my existing java programs with it.  I think my 
CoffeeLink nntp server would see a big benefit from deployment as native 
code.  I just don't think it's a good alternative to the C++ code base 
for a tool like spamprobe that needs to run frequently and 
simultaneously under many user accounts.

I was really hoping to see compiled java running within 80% or so of 
compiled C++ code so I could have the productivity benefits of java 
without much of a performance hit.  I'd have danced around the room if 
native java had been faster. :-)

Now I'm thinking of porting over some of the code in CoffeeLink into 
spamprobe.  I have some nice code for config files there and even a 
little b-tree package that I'm anxious to run side by side with BDB/PBL 
to see how it compares.  It's probably slower but it has full 
transaction semantics which can be nice for something like spamprobe to 
avoid database corruption in the event of a system crash.

I'm starting to think that a multi-user persistent daemon version of SP 
might be worth having too.  It could talk to a SQL database and keep 
enough info in the database to make periodic training updates faster 
(i.e. store lists of messages and their tokens in the database to avoid 
having to reparse them).

All the best,
++Brian

-----BEGIN PGP SIGNATURE-----
Version: PGP Desktop 9.0.2 (Build 2425)

iQEVAwUBQ3uaTjxRyEoJfXIFAQg0rAf/Tu75mFAfFVFydJpdrFNUQrjh9KTb5Cj7
z3fF54E9wZVmBUcz7DLCEHA7qWImvnm7uUmJAL7kwmzAnyAMBCNiGAz+Itw+XufS
B8lYMQOaQaDUGw8abQatQjzcfbMLYJqPyAaFGgFbVmDtEQQ6BClppNuCaBHGgWch
lcyK9V+mwy5mFSuRbChGlJ/ybRBvGi/nDzsxIJHX9zK6CrYVMrG+GFv0Yxg7G7cM
O7SJmVMyClUOx/6Q7CDyAxMhf/K14PaRh5nDAOrV7JGU6OXNgh0Ifi+QE6+lbMFj
hfxFsXUiZDFyTTt37CZS8hNWXyEvaR+FrhL6VOqfn1sS7kDEdQ1nHA==
=DDKw
-----END PGP SIGNATURE-----


-------------------------------------------------------
This SF.Net email is sponsored by the JBoss Inc.  Get Certified Today
Register for a JBoss Training Course.  Free Certification Exam
for All Training Attendees Through End of 2005. For more info visit:
http://ads.osdn.com/?ad_id=7628&alloc_id=16845&op=click