Re: SmbTransport thread accumulation issues

I believe some Java 6 NIO calls use epoll on Linux (if the underlying o/s
has this support) and select on Windows (which is pretty reliable in my
experience)

http://download.oracle.com/javase/6/docs/technotes/guides/io/enhancements.html

There's a far amount of debate about non-blocking IO versus thread based
blocking IO kicking around the internet - it seems the current consensus is
NIO for massive scalability and thread based for speed. A lot of the
benchmarks are very dependant on the underlying o/s

C

Am 18.10.2011 20:59, schrieb Michael B Allen:
> > BTW: I tried NIO once. Once. It is a crumby implementation of
> > select(2) from UNIX (which was always considered by many to be a
> > rather crude design in itself). NIO should be called "NoIO". I had
> > trouble just getting it to work reliably. And note that NIO would not
> > really help that much. It would only help in this particular case
> > where you want to communicate with 100 servers in parallel. Otherwise,
> > JCIFS is actually extremely efficient. The CIFS protocol allows for
> > multiplexing requests. Meaning you can send request A and B and get
> > the response for B before A. And JCIFS fully utilizes this. JCIFS can
> > touch every file on a workstation in under 10 seconds if the server
> > cache is hot. If the non-blocking implementation in Java were better I
> > might consider using it. But it's not.
> >
> > Mike
> >
> Hi Mike,
>
> you are right if you say that selector implementation (introduced
> in java 1.4) isn't the easiest part to implement against. And
> can lead very quick to the meaning that the socket-part of the
> NIO in Java 1.4 should be nearly named "NoIO". But i have done
> this part of porting on jcifs (in my private local git-branch) and
> even if i am not finished with it, it actually works quite good.
> And yes i hate that selector implementation too.
>
> Java 7 introduces new classes for Asynchronous Network IO
> (AsynchronousSocketChannel) with CompletionHandling. This
> really makes the task easier to use asyn-network-io. But i think
> jcifs is implemented against the java5 feature-set. Am i right
> with the latter?
>
> -- Sebastian
>
>