[Biojava-l] Java Resource Management [a semi troll...]

[Thomas Down]

On Sun, Feb 09, 2003 at 04:56:28PM -0500, Dave Keller wrote:
> 
> >This is why I think i have alot of reluctance to move to Java; C is by far
> >my prefered strongly-typed, "non-scripting" language. But I am a
> >dinosaur...
> > 
> >
> Nothing wrong with that, C is an excellent language and there is still 
> plenty of C development going on.  If C is the right tool for the job, 
> then by all means use it.
> 
> Speaking of C, I was wondering if anybody has done any experimenting 
> with using C/C++ and JNI for computational and/or resource intensive 
> tasks.  Since BioJava is based on interfaces, it seems you could replace 
> sequence implementations and objects that perform calculations on 
> sequences with C implementations.

I'll put in a mild bite on this one.  Many years ago, I wrote
a little tool which did some naive (but rather computationally
intensive -- or at least it felt that way on the hardware
available at the time ;-).  It had an AWT frontend, a main loop
doing the actual calculations, and some graphing routines grafted
on the back.  You know the kind of thing...

Anyway, version 1 was pure Java all the way through.  It was
a little slow, so I re-wrote the whole of the main loop in C,
and called it through JNI.  I got something like a 3-4 times
speedup at the time, which, right then, was useful to me.
However, there are a few points to note about this:

  - At the time I was using the very first (probably counts as
    an alpha) port of JDK1.2 to Linux.  This included the
    dearly departed `sunwjit' compiler, which was *nowhere near*
    as fast as the current VMs (especially Hotspot server).  I'm
    actually tempted now to dig it all out, see if I can still
    get the JNI bits to compile, and post new benchmarks.  Maybe
    tommorow.

  - You only got speed benefits if you shunted the whole loop
    across to C.  Native-coding the body didn't help at all,
    and maybe even made things worse (I can't remember now).
    But JNI calls were, at the time, seriously expensive.

The cost of Java -> C calling makes me suspect that naively
making native implementations of the BioJava stuff is unlikely
to be worthwhile in the vast majority of cases.  That said,
I believe that JNI costs are substantially lower in JDK1.4.x,
so it's something which might be worth reconsidering now.
Does anyone have *recent* JNI experience, and want to comment
on this?

       Thomas.