1. This forum section is a read-only archive which contains old newsgroup posts. If you wish to post a query, please do so in one of our main forum sections (here). This way you will get a faster, better response from the members on Motherboard Point.

Java performance issue on a M5000 (Solaris 10u9) - known issue ?

Discussion in 'Sun Hardware' started by Frank Zimmer, Apr 22, 2011.

  1. Frank Zimmer

    Frank Zimmer Guest

    Hi,

    we are currently setting up some Oracle Fusion stuff and during that we
    encountered some performance issues, slow start, slow requests etc.
    Because of that we started some Specjvm benchmarks and detected that a
    simple HP Server with 2 quad core Intel CPU (5520) is (much) faster than
    our M5000 with 8 Sparc VII (128 RB RAM).

    Does anyone else have some comparable experiences, ideas or tips.
    We have tuned the JVM with some settings from official benchmarks
    (adopted to our system) - so please do not ask for Java tuning.

    We feel a bit uncomfortable, it is a kind of feeling we have that
    something may be wrong with the box.

    Thanks
    Frank
     
    Frank Zimmer, Apr 22, 2011
    #1
    1. Advertisements

  2. The SPARC CPU deliver about 50% of current Intel CPUs for single
    threaded apps. See SPEC CPU2006 benchmarks.
    Since you are setting up the application you probably did more or less
    single user tests so far.
    But for a high number of concurrent users you have a very reliable and
    stable system with a good overall throughput as long as the application
    is designed for this.
     
    Frank Langelage, Apr 22, 2011
    #2
    1. Advertisements

  3. how many threads are you running in your tests?
     
    Cydrome Leader, Apr 22, 2011
    #3
  4. Frank Zimmer

    Rick Jones Guest

    A caveat about SPECint2006 or SPECint_base2006 (or their "fp" cousins)
    and single-threaded performance. For quite some time (since the
    beginning of SPECcpu2006 I suspect), SPEC has allowed such "speed"
    results to use auto-parallelization features of the compiler(s). As
    such, unless you look at the full disclosure report and see that it
    says "Auto Parallel: No" you must not ass-u-me the result represents
    single anything performance - neither single-thread, nor even single
    core or single processor. (Ok, *perhaps* single system, but then one
    cannot ass-u-me that of "rate" results...without checking the FDR)

    If the FDR for the speed run says "Auto Parallel: Yes" then you have
    to start looking more closely - see if an auto-parallelizing option
    was used in base for one or more of the programming languages, and/or
    if one was used on one or more sub-benchmarks in peak.

    For the last several years, 99 times out of 10 (*) the "speed" metrics
    you will find on www.spec.org have used auto parallelization. The
    truly single-threaded results are much more rare. Until you can
    examine the FDRs you should assume a given speed result is *not*
    single-threaded.

    happy benchmarking,

    rick jones

    (*) yes that does say 10 - a saying I learned years ago in high-school
    from a sports coach (Thank you Coach McNair, wherever you are :), and
    means "virtually all the time."
     
    Rick Jones, Apr 23, 2011
    #4
  5. Do make sure you are running the latest JVM - there have been lots
    of performance enhancements going in over recent updates. In S10U9,
    which JVM you are running by default depends on if you upgraded (1.5)
    or if you did a fresh install (1.6), but in either case, it may be
    worth fetching the latest update now.
     
    Andrew Gabriel, Apr 23, 2011
    #5
  6. Frank Zimmer

    ITguy Guest

    we are currently setting up some Oracle Fusion stuff and during that we
    What's worse is that the M5000 is easily 10X and likely over 20X the
    cost of your commodity box.
    My experience has been the same - X86 chip performance has been better
    than SPARC some time. The simple answer is to migrate off SPARC.
    Go back to the reasons the M5000 was selected as the deployment
    platform in the first place. Are they still valid?
     
    ITguy, Apr 24, 2011
    #6
  7. Frank Zimmer

    Frank Zimmer Guest

    We run 8 threads on the M5000 and 8 on the Intel box, so we compare 8
    cores against 8 cores.
    Of course, if we run the Specjvm with the default values, on the M5000
    there will be 128 threads running against 16 on the intel box, but this
    would not be a fair thing.

    Frank
     
    Frank Zimmer, Apr 26, 2011
    #7
  8. Frank Zimmer

    Frank Zimmer Guest

    JDK 1.6.24, Solaris 10u9

    so nothing more to do
     
    Frank Zimmer, Apr 26, 2011
    #8
  9. Frank Zimmer

    Frank Zimmer Guest

    Yes, they are.
    I only wanted to get some experiences in order to avoid any
    "misconfiguration" or maybe a hardware fault (slow RAM, whatwever) on
    the Sparc box.
     
    Frank Zimmer, Apr 26, 2011
    #9
  10. Frank Zimmer

    odk Guest

    W dniu 2011-04-26 10:29, Frank Zimmer pisze:
    Look using dtrace at java process calls. If there is a lot of garbage
    collection calls, You found your issue.
     
    odk, Apr 26, 2011
    #10
  11. Frank Zimmer

    Frank Zimmer Guest


    No significant GC's, less than 1 % of overall cpu time.
     
    Frank Zimmer, Apr 26, 2011
    #11
  12. Frank Zimmer

    Jim Leonard Guest

    That's kind-of the point. Sparc's strength is in multithreaded
    processing, not single-core performance.

    If you're going to limit the sparc to a certain number of hardware
    threads, then maybe you should lower the clock speed of the Intel
    chips to match that of the M5000? To make it more "fair"? ;-)
     
    Jim Leonard, Apr 26, 2011
    #12
  13. Frank Zimmer

    Frank Zimmer Guest


    I know ;-)

    but in this case clock speeds are nearly the same.
    An of course I also know that there are differences in multi threading.

    But also with that in mind the 8 cores of the Intel box outperform the 8
    cores of the SPARC, by more than 30 %, and this is where I wanted to
    know if this is "normal" or caused by a potential misconfiguration of
    our M5000, or if there is a "magic" kernel parameter, OpenBoot
    configuration which leads to a slower memory access or whatwever.

    Not only for the Java benchmark, also for some Oracle databases we have
    the feeling !!!! (no real indicators) that the whole Server is a bit
    slow, and here we want to get other experiences.

    And please don't bother with certain JVM tunings etc.
    I am working with Sun Servers since now more than 15 years and i believe
    (d) in them. I only want to have some reasons why we feel this
    performance issue.

    Regards
    Frank
     
    Frank Zimmer, Apr 26, 2011
    #13
  14. I don't think running 8 theads on any box is a test of anything. Give the
    machine a real load and see how they scale as the load goes up. It's never
    linear and can be interesting.
     
    Cydrome Leader, Apr 26, 2011
    #14
    1. Advertisements

Ask a Question

Want to reply to this thread or ask your own question?

You'll need to choose a username for the site, which only take a couple of moments (here). After that, you can post your question and our members will help you out.