Low CPU utilization on Solaris

Question

Hi all.
We've recently been performance tuning our java application running
inside of an Application Server with Java 1.3.1 Hotspot -server. We've
begun to notice some odd trends and were curious if anyone else out
there has seen similiar things.
Performance numbers show that our server runs twice
as fast on Intel with Win2K than on an Ultra60 with Solaris 2.8.
Here's the hardware information:
Intel -> 2 processors (32bit) at 867 MHz and 2 Gig RAM
Solaris -> 2 processors (64bit) at 450 MHz and 2 Gig RAM.
Throughput for most use cases in a low number of threads is twice as
fast on Intel. The only exception is some of our use-cases that are
heavily dependent on a stored procedure which runs twice as fast on
Solaris. The database (oracle 8i) and the app server run on the same
machine in these tests.
There should minor (or no) network traffic. GC does not seem to be an
issue. We set the max heap at 1024 MG. We tried the various solaris
threading models as recommended, but they have accomplished little.
It is possible our Solaris machine is not configured properly in some
way.
My question (after all that ...) is whether this seems normal to
anyone? Should throughput be higher since the processors are faster on
the wIntel box? Does the fact that the solaris processors are 64bit
have any benefit?
We have also run the HeapTest recommended on this site on both
machines. We found that the memory test performs twice as fast on
solaris, but the CPU test performs 4 times as slow on solaris. The
"joint" test performs twice as slow on solaris. Does this imply bad
things about our solaris configuration? Or is this a normal result?
Another big difference is between Solaris and Win2K in these runs is
that CPU Utilization is low on solaris (20-30%) while its much higher
on Win2K (60-70%)
[both machines are 2 processor and the tests are "primarily" single
threaded at
this stage]. I would except the solaris CPU utilization to be around
50% as well. Any ideas why it isn't?

Answer 1

Hi,
I recently went down this path and wound up coming to the realization that the
cpu's are almost neck and neck per cycle when running my Java app. Let me qualify
this a little more (400mhz Sparc II cpu vs 500mhz Intel cpu) under similar load
running the same test gave me similar results. It wasn't as huge difference in
performance as I was expecting.
My theory is given the scalability of the SPARC architecture, more chips==more
performance with less hardware, whereas the Wintel boxes are cheaper, but in order
to get scaling, the underlying hardware comes into question. (how many wintel
boxes to cluster, co-locate, manage, etc…)
From what little I've found out when running tests against our Solaris 8 (E-250's)
400mhz UltraSparc 2's is that it appears that the CPU performance in a lightly
threaded environment is almost 1 cycle / 1 cycle (SPARC to Intel). I don't think
the 64 bit SPARC architecture will buy you anything for java 1.3.1, but if your
application has some huge memory requirements, then using 1.4.0(when BEA supports
it) should be beneficial (check out http://java.sun.com/j2se/1.4/performance.guide.html).
If your application is running only a few threads, tying the threads to the LWP
kernel processes probably won't gain you much. I noticed that it decreased performance
for a test with only a few threads.
I can't give you a good reason as to why your Solaris CPU utilization is so low,
you may want to try getting a copy of Jprobe and profiling Weblogic and your application
to see where your bottlenecks are. I was able to do this with our product, and
found some nasty little performance bugs, but even with that our CPU utilization
was around 98% on a single and 50% on a dual.
Also, take a look at iostat / vmstat and see if your system is bottlenecking doing
io operations. I kept a background process of vmstat to a log and then looked
at it after my test and saw that my cpu was constantly pegged out (doing a lot
of context switching), but that it wasn't doing a whole lot of page faults
(had enough memory).
If you're doing a lot of serialization, that could explain slow performance as
well.
I did follow a suggestion on this board of running my test several times with
the optimizer (-server) and it boosted performance on each iteration until a plateau
on or about the 3rd test.
If you're running Oracle or another RDBMS on your Solaris machine you should see
a pretty decent performance benchmark against NT as these types of applications
are more geared toward the SPARC architecture. From what I've seen running Oracle
on Solaris is pretty darn fast when compared to Intel.
I know that I tried a lot of different tweaks on my Solaris configuration (tcp
buffer size, etc/system parameters for file descriptors, etc.) I even got to the
point where I wanted
to see how WebLogic was handling the Nagle algorithm as far as it's POSIX muxer
was concerned and ran a little test to see how they were setting the sockets (setTcpNoDelay(Boolean)
on java.net.Socket). They're disabling the Nagle algorithm so that wasn't an
issue sigh. My best advice would be to profile your application and see where
the bottlenecks are, you might be able to increase performance, but I'm not too
sure. I also checked out www.spec.org and saw some of their benchmarks that
coincide with our findings.
Best of luck to you and I hope this helps :)
Andy
[email protected] (feanor73) wrote:
Hi all.
We've recently been performance tuning our java application running
inside of an Application Server with Java 1.3.1 Hotspot -server. We've
begun to notice some odd trends and were curious if anyone else out
there has seen similiar things.
Performance numbers show that our server runs twice
as fast on Intel with Win2K than on an Ultra60 with Solaris 2.8.
Here's the hardware information:
Intel -> 2 processors (32bit) at 867 MHz and 2 Gig RAM
Solaris -> 2 processors (64bit) at 450 MHz and 2 Gig RAM.
Throughput for most use cases in a low number of threads is twice as
fast on Intel. The only exception is some of our use-cases that are
heavily dependent on a stored procedure which runs twice as fast on
Solaris. The database (oracle 8i) and the app server run on the same
machine in these tests.
There should minor (or no) network traffic. GC does not seem to be an
issue. We set the max heap at 1024 MG. We tried the various solaris
threading models as recommended, but they have accomplished little.
It is possible our Solaris machine is not configured properly in some
way.
My question (after all that ...) is whether this seems normal to
anyone? Should throughput be higher since the processors are faster on
the wIntel box? Does the fact that the solaris processors are 64bit
have any benefit?
We have also run the HeapTest recommended on this site on both
machines. We found that the memory test performs twice as fast on
solaris, but the CPU test performs 4 times as slow on solaris. The
"joint" test performs twice as slow on solaris. Does this imply bad
things about our solaris configuration? Or is this a normal result?
Another big difference is between Solaris and Win2K in these runs is
that CPU Utilization is low on solaris (20-30%) while its much higher
on Win2K (60-70%)
[both machines are 2 processor and the tests are "primarily" single
threaded at
this stage]. I would except the solaris CPU utilization to be around
50% as well. Any ideas why it isn't?