Christopher Smith wrote:
> On 31 May 2001 06:45:08 +1000, Jesus M. Salvo Jr. wrote:
> > > 4) Use JNI to use Linux's various asynch I/O API's.
> > Option 4) is how BEA WebLogic Server does it, ( I think ). They have this
> > libmuxer.so ( which is also available for Solaris -- dont know why when JVM
> > for Solaris makes use of solaris native threads ) which enables native I/O,
> > or the so-called "performance pack".
>
> Yes, many of the various application servers employ various strategies
> for this. Unfortunately I am unaware of any of them who have taken the
> time to do this for Linux.
>
> > Speaking of 4), does anyway have a ready-made ( cut&paste ) C code that one
> > can compile into a shared library and have all I/O made through this library
> > via JNI?
>
> I'll be demoing a library that does this with SGI's KAIO patch at the
> presentation (presuming I can iron out this nasty sigtimedwait problem).
>
> > And how does JDK 1.4 affect the performance on Linux, given that 1.4 was
> > better I/O support overall, in particular non-blocking I/O.
>
> We're benchmarking that right now, but it doesn't take a genius to
> realize the impact is pretty spectacular. All you need to do is compare
> Apache's performance with polling I/O based web servers on Linux to get
> a good idea of the potential performance gains.
>
> The NIO API is a little annoying to me though in that it's low enough
> level that it exposing the polling approach to I/O. With the overhead of
> JNI, it's much nicer to use an asynch I/O approach for most Java
> applications.
Chris,
Thanks for providing this pre-session back and forth. Although I'm excited by
the
prospects and possibility of say option number 4 over the long haul, I've got to
be concerned as much with stability as scalability. After all, I've got
to serve actually existing applications in production today. ;-)
So, the green threads or the lots of boxes w/tweaking of the kernel have been the
options
of choice for me and my colleagues. Some people claim that green threads are not
at all
an option when scaling is your objective, so I'm heartened to see that even as
you are
experimenting with SGI's KAIO patch, you've kept green threads as an option.
I assume you are going to address the tradeoffs of the various options in your
presentation?
I'm interested, you might have gathered, in practical, and as stable as possible,
options for
scaling Java on Linux.
Thanks again,
Ed
----------------------------------------------------------------------
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
