Hi Matt,

I did some processor scalability tests with VolanoMark back in August,
1998, testing Microsoft's and Sun's Java virtual machines on one through
four processors (both Intel and SPARC).  The results are under the
section called "The hardware solution" in my JavaWorld article:

  "Which Java VM scales best?"
  http://www.javaworld.com/javaworld/jw-08-1998/jw-08-volanomark_p.html

I don't test processor scalability anymore because, for me, it's not the
problem.  (It used to be a problem back in 1997 when Sun's Java VM went
two times *slower* on Solaris when you added a second processor.)  These
days, for pure Java servers with lots of dedicated long-lasting socket
connections, the problem is network scalability:

  VolanoMark 2.1.2 Network Scalability (Line Plot)
  http://www.volano.com/report/network2.gif

Note that the limits on the number of concurrent connection exist no
matter how many processors or how much memory you throw at the problem. 
Only Sun's 1.3.0 HotSpot Server VM on Solaris had problems under 6,000
connections because of real memory (320 MB).  All of the other problems
are due to architectural limits or bugs.  These limits used to be due to
file descriptors and thread handling, but now it's just the thread
handling.  The details of each failure are shown if you click on the
chart to get this page:

  VolanoMark 2.1.2 Network Scalability (Details)
  http://www.volano.com/report/network.txt

The Windows VMs using a one-to-one mapping of Java to native threads
allocate a stack reserve of 256 KB per thread and fail (with an
OutOfMemoryError) when the reserved storage exhausts the 2 GB address
space.  The Linux VMs using a one-to-one mapping fail when they hit the
maximum number of Linux tasks or limits in the Linux threads library. 
The other failures are due to bugs.

There are, of course, applications which need good processor
scalability.  But as far as VolanoMark and our VolanoChat product are
concerned, what good is processor scalability, or even superb
performance in general, when you can only connect 422 people?

John Neffenger

Matt Peterson wrote:
> 
> John,
> 
> Thanks for sending the results of the Volano benchmark to the list.  It
> was interesting to see the performance comparisons.  However, I was
> disappointed to see that SMP hardware was not used in the test.  Is
> there a chance that we will see published numbers that show results on
> SMP hardware?  It would definitely be interesting to those that are
> interested in seeing how native thread Java scales in comparison to
> those JVMs that do not use kernel (system mode) thread libraries for
> their thread implementations.  It would also suggest the suitability of
> various platforms (Linux vs Windows vs Solaris) for implementation of a
> fast scalable JVM.
> 
> --
> Matthew Peterson
> Sr. Software Engineer
> Caldera Systems, Inc
> [EMAIL PROTECTED]


----------------------------------------------------------------------
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]

Reply via email to