Comments for Ismael Juma http://blog.juma.me.uk Thu, 14 Jul 2011 17:33:59 +0000 hourly 1 http://wordpress.com/ Comment on 32-bit or 64-bit JVM? How about a Hybrid? by Ismael Juma http://blog.juma.me.uk/2008/10/14/32-bit-or-64-bit-jvm-how-about-a-hybrid/#comment-528 Thu, 14 Jul 2011 17:33:59 +0000 http://ijuma.wordpress.com/?p=32#comment-528 Sorry Usman, the information you provided is not enough to help you. I suggest you ask on stackoverflow or a site like it.

]]> Comment on 32-bit or 64-bit JVM? How about a Hybrid? by Ismael Juma http://blog.juma.me.uk/2008/10/14/32-bit-or-64-bit-jvm-how-about-a-hybrid/#comment-527 Thu, 14 Jul 2011 17:32:58 +0000 http://ijuma.wordpress.com/?p=32#comment-527 Hi Jumar. Yes, it is enabled automatically when relevant if you’re using update 23 or later.

]]> Comment on 32-bit or 64-bit JVM? How about a Hybrid? by Jumar http://blog.juma.me.uk/2008/10/14/32-bit-or-64-bit-jvm-how-about-a-hybrid/#comment-526 Thu, 14 Jul 2011 17:26:17 +0000 http://ijuma.wordpress.com/?p=32#comment-526 To clarify your last comment:
“This feature is enabled by default since JDK 6 Update 23″.
This means that as of jdk6u23, the option -XX:+UseCompressedOops is enabled automatically when relevent? Or just that this option is supported if specified as command line?

]]> Comment on 32-bit or 64-bit JVM? How about a Hybrid? by Stefan’s Kugel » Blog Archive » [Java] Faster with -XX:+UseCompressedOops or upgrade to jdk1.6.0_23 http://blog.juma.me.uk/2008/10/14/32-bit-or-64-bit-jvm-how-about-a-hybrid/#comment-525 Sat, 18 Jun 2011 16:56:22 +0000 http://ijuma.wordpress.com/?p=32#comment-525 [...] the old 4GB limit, through some optimizations. There are plenty of good post with the details, e.g. http://blog.juma.me.uk/2008/10/14/32-bit-or-64-bit-jvm-how-about-a-hybrid/. Java 1.6.0_22 is the last version which has this option disabled by default. Since 1.6.0_23 it is [...]

]]> Comment on 32-bit or 64-bit JVM? How about a Hybrid? by Usman Aleem http://blog.juma.me.uk/2008/10/14/32-bit-or-64-bit-jvm-how-about-a-hybrid/#comment-524 Mon, 06 Jun 2011 19:27:16 +0000 http://ijuma.wordpress.com/?p=32#comment-524 Here is my problem i am working on a project which uses javaws(32 bit jre) and my application build on some wrappers C++. when i try to run it with 1024M it is working fine on 64bit machine windows7 but when i try to run it with 1536M. outof memory and very strange behavious observed like one program working fine and some time it even gives me null pointer exception on calling simple JFileChooser dialog etc. and then that application will never works unless restart my application or just hangs.

]]> Comment on Performance of FastMath from Commons Math by Ismael Juma http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-390 Sat, 26 Feb 2011 14:44:38 +0000 http://blog.juma.me.uk/?p=184#comment-390 Hi Rudiger,

Thanks for the link, it looks interesting. I’ll add it to my list of things to try. :)

Best,
Ismael

]]> Comment on Performance of FastMath from Commons Math by Rüdiger Keller http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-389 Sat, 26 Feb 2011 09:53:44 +0000 http://blog.juma.me.uk/?p=184#comment-389 I recently stumbled upon a “replacement” for java.lang.Math called Jafama. It claims to have far superior performance but does not provide full accuracy. Quoted from the project page: “usually about 2-4 (up to 15) times faster, with about 1e-15 accuracy”. Although I must admit I have never used it. But perhaps benchmarking it would provide some interesting additions to your graphs. I at least would be very interested in performance numbers on this.

http://sourceforge.net/projects/jafama/

]]> Comment on Performance of FastMath from Commons Math by Ismael Juma http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-384 Fri, 25 Feb 2011 10:58:57 +0000 http://blog.juma.me.uk/?p=184#comment-384 Hi Kirk,

If you look at the test, you’ll see that the numbers are not small, they are just converted at the end. Also, the scores were very stable across iterations. Having said that, it would be nice to have variance scores and I welcome you or Aleksey to submit the changes to the code (it’s all available after all). I’ll add the data to the post if that happens.

About the lack of confidence in the numbers, that’s good. You should not trust performance numbers in random blogs (independently of whether they contain variance calculations or not). Instead you should perform tests with your own hardware and application (even more true for math operations). Random blogs can serve at best as the starting point of an investigation. At least that’s the approach I take.

Best,
Ismael

]]> Comment on 32-bit or 64-bit JVM? How about a Hybrid? by Ismael Juma http://blog.juma.me.uk/2008/10/14/32-bit-or-64-bit-jvm-how-about-a-hybrid/#comment-383 Fri, 25 Feb 2011 10:50:38 +0000 http://ijuma.wordpress.com/?p=32#comment-383 The feature I am talking about in the post and in the quoted sentence is -XX:+UseCompressedOops (not -XX:+ConcMarkSweepGC).

]]> Comment on Performance of FastMath from Commons Math by Kirk Pepperdine http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-382 Thu, 24 Feb 2011 20:33:53 +0000 http://blog.juma.me.uk/?p=184#comment-382 Aleksey’s comments on reporting are spot on. Adding a variance calculation should be automatic. And, to say that values are different, you need further statistical testing. Without the variance calculation it’s hard for me to have confidence that the small numbers are the same or significantly different.

]]> Comment on Performance of FastMath from Commons Math by Ismael Juma http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-377 Thu, 24 Feb 2011 09:13:33 +0000 http://blog.juma.me.uk/?p=184#comment-377 Hey Tatu,

That was the motivation indeed.

Good point about “lower is better”. I’ve updated the post to mention this.

Best,
Ismael

]]> Comment on 32-bit or 64-bit JVM? How about a Hybrid? by phoenix http://blog.juma.me.uk/2008/10/14/32-bit-or-64-bit-jvm-how-about-a-hybrid/#comment-373 Thu, 24 Feb 2011 02:59:10 +0000 http://ijuma.wordpress.com/?p=32#comment-373 “This feature is enabled by default since JDK 6 Update 23″.
Is this mean that we don’t need add -X:+UseConcMarkSweepGC after JDK 6 Update 23? And how can we check if the JDK enabled this default? Thanks!

]]> Comment on Performance of FastMath from Commons Math by Tatu Saloranta http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-371 Wed, 23 Feb 2011 21:18:03 +0000 http://blog.juma.me.uk/?p=184#comment-371 Good stuff: I think it is very important to have independent verification of claims made by lib authors, as well as check how performance evolves over time with newer JDKs. Too often data from old JVMs is used as gospel.

One very minor suggestion — it would be good to mention that graphs represent time in milliseconds (i.e. “shorter is better”). This because there are tests where opposite is through (value is throughput).

]]> Comment on Performance of FastMath from Commons Math by Aleksey Shipilev http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-366 Wed, 23 Feb 2011 13:57:39 +0000 http://blog.juma.me.uk/?p=184#comment-366 Ok, I’m convinced then.

The only thing separating your work from nice performance analysis is having confidence intervals or any other sort of error estimation for the data.

Thanks and over,
Aleksey.

]]> Comment on Performance of FastMath from Commons Math by Ismael Juma http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-365 Wed, 23 Feb 2011 13:54:50 +0000 http://blog.juma.me.uk/?p=184#comment-365 Patrick, I am aware of those tools and I’d use one of them if I were writing the tests from scratch. In this case, I think Aleksey’s concerns are unfounded and I’m happy with the results as they are.

Best,
Ismael

]]> Comment on Performance of FastMath from Commons Math by Patrick http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-364 Wed, 23 Feb 2011 13:51:46 +0000 http://blog.juma.me.uk/?p=184#comment-364 0.02 on benchmarking: this is why tools like Japex and Caliper were created. Perhaps using one of those would alleviate some of the concern Alexis has?

]]> Comment on Performance of FastMath from Commons Math by Ismael Juma http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-363 Wed, 23 Feb 2011 13:51:46 +0000 http://blog.juma.me.uk/?p=184#comment-363 I don’t agree that I measure time in interpreted mode for the time used in the charts and the tables. To be clear, I am only using the scores from the 10th iteration and -XX:+PrintCompilation is silent then.

Feel free to run the tests yourself and check. In my machines the scores are very stable, which is a good sign.

]]> Comment on Performance of FastMath from Commons Math by Aleksey Shipilev http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-362 Wed, 23 Feb 2011 13:39:18 +0000 http://blog.juma.me.uk/?p=184#comment-362 Ismael, the data and test sure look more sane now.

However, I’m still not convinced there’s enough time for JVM warmup. You still measure time in interpreted mode, then measure overhead of OSR, then measure the compiled code performance. To add more, since all three tests are in the same method, you may end up in situation where StrictMath (first in your method) is running in interpret mode, exceeding compilation threshold, forcing entire method to recompile, thus given advantage for Math and FastMath executing in pure compiled mode afterwards.

I see this is mitigated by having 10 cycles of entire benchmark run in test(). Is that enough? Is -XX:+PrintCompilation silent when you actually start measuring?

Note, this also implies you have 10 report() calls per particular test, hence 10 samples, in which first samples are warmup garbage. I wonder if new charts account for last cycle only?

Also, while you have these 10 samples, can you actually compute the confidence intervals for your data? Some of percieved differences might be gone then.

Sorry for being so obnoxious.

]]> Comment on Performance of FastMath from Commons Math by Mark Wielaard http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-361 Wed, 23 Feb 2011 13:29:47 +0000 http://blog.juma.me.uk/?p=184#comment-361 I should have added that having StrictMath “pure java” does rely on the VM implementing the strictfp modifier correctly. Something which is sometimes hard to get right. IKVM for example doesn’t.
http://weblog.ikvm.net/CommentView.aspx?guid=6d657cd6-58ac-4a6a-8be9-c3fba167f531

]]> Comment on Performance of FastMath from Commons Math by Ismael Juma http://blog.juma.me.uk/2011/02/23/performance-of-fastmath-from-commons-math/#comment-360 Wed, 23 Feb 2011 13:24:46 +0000 http://blog.juma.me.uk/?p=184#comment-360 Hey Mark,

Interesting. It would be good to see how the pure Java implementation performs.

Best,
Ismael

]]>