> Comments would be appreciated. Be aware that it is sketchy and unfinished.
I've got some couple comments for you. I've been working on this
problem (multiple processes in a JVM) at the University of Utah for a
while now. Our research group has a paper which was just accepted to
the April USENIX on just this topic. The paper tries to provide a
solid list of the issues involved and choices available for Java
process models. You can get it at [1]. My Master's thesis [2] also
contains a lot of relevant info, but its not nearly as concise or
readable. :) There is also other work in java processes here at Utah
(GVM [1], KaffeOS [3]) and at Cornell (JKernel [4], Luna [5]). And,
some older work at Princeton ([6]).
One nit on your definition of class version. Two classes are the same
only if the bytecode is identical and if all the referenced classes
are identical. If I load a class BigButton into two different class
loaders, and in each classloader BigButton's superclass Widget is
different, the the two instances of BigButton are different. The JVM
spec makes a coarser restriction: classes loaded into separate
classloaders are by definition different, even if all the classes are
actually the same.
Another issue is the tension between sharing objects and total
reclaimation of memory. If a process is terminated, but some objects
it allocated are sitting in a global table, what do you do? What if
that object has pointers all of the data area of the just-terminated
process? There are systems (such as Luna) that can reclaim the
object, though Luna makes changes to Java's type system and requires
some pretty hefty optimizations to avoid significant run-time
penalties. Other systems, such as KaffeOS strictly limit the types
of such shared objects, thus limiting the damage that can be done.
Other systems (such as the one I'm working on now) disallow shared
objects at the process level (the "system" can share objects but it is
trusted to clean them up).
If you're going for a UNIX-like process model, I suggest you start by
disallowing sharing (copy objects across process and process/kernel
boundaries). I get the feeling this is the sort of process model you
are heading towards.
Good luck.
-Pat
----- ----- ---- --- --- -- - - - - -
Pat Tullmann [EMAIL PROTECTED]
This signature witticism intentionally left blank.
[1] http://www.cs.utah.edu/flux/papers/javaos-tr98015-abs.html
[2] http://www.cs.utah.edu/~tullmann/thesis/thesis.html
[3] Godmar doesn't have a good link yet, http://www.cs.utah.edu/~gback/
[4] http://www.cs.cornell.edu/slk/JKernel/Default.html
[5] http://www.cs.cornell.edu/Info/People/hawblitz/PLDI2000-submit/luna-99-11-13.ps
[6] http://www.cs.princeton.edu/sip/pub/icdcs.html
_______________________________________________
Kernel maillist - [EMAIL PROTECTED]
http://jos.org/mailman/listinfo/kernel
