Re: Review request for 6977034 Thread.getState() very slow

Sat, 04 Dec 2010 04:23:54 -0800

Hi Mandy,

This test:

        if ((threadStatus & JVMTI_THREAD_STATE_RUNNABLE) == 1) {

is always false, since JVMTI_THREAD_STATE_RUNNABLE is 4. (NetBeans 7.0 helpfully flags this; I'm not sure if earlier versions do.)

But, once corrected, I think you could use this idea further to write a much simpler and faster method, on these lines:

    public static Thread.State toThreadState(int threadStatus) {
        if ((threadStatus & JVMTI_THREAD_STATE_RUNNABLE) != 0) {
            return RUNNABLE;
        } else if ((threadStatus & JVMTI_THREAD_STATE_BLOCKED_ON_MONITOR_ENTER) != 0) {
            return BLOCKED;
        } else if ((threadStatus & JVMTI_THREAD_STATE_WAITING_WITH_TIMEOUT) != 0) {
            return TIMED_WAITING;
        } else if ((threadStatus & JVMTI_THREAD_STATE_WAITING_INDEFINITELY) != 0) {
            return WAITING;
        } else if ((threadStatus & JVMTI_THREAD_STATE_TERMINATED) != 0) {
            return TERMINATED;
        } else {
            return NEW;
        }
    }

You could tweak the order of the tests based on what might be the relative frequency of the different states but it probably isn't worth it.

Regards,
Éamonn

On 3/12/10 11:52 PM, Mandy Chung wrote:
Fix for 6977034: Thread.getState() very slow

Webrev at:
   http://cr.openjdk.java.net/~mchung/6977034/webrev.00/

This is an improvement to map a Thread's threadStatus field to Thread.State.  The VM updates the Thread.threadStatus field directly at state transition with the value as defined in JVM TI [1].  The java.lang.Thread.getState() implementation can directly access the threadStatus value and do a direct lookup from an array of Thread.State.  The threadStatus value is a bit vector and we would have to create an array of a minimum of 1061 (0x425) elements to do direct mapping.   I took the approach to use the first highest order bit set to 1 in the masked threadStatus value as the index to the Thread.State element and only caches 32 elements (could be fewer).  I wrote a micro-benchmark measuring the Thread.getState of a thread in different state that shows 1.7X to 6X speedup (see below).  There is possibly some issue with my micro-benchmark that I didn't observe the 14X speed up as Doug did in his experiment.  However, I'd like to get this reviewed and pushed to the repository so that anyone can do more experiment on the performance measurement.

Thanks
Mandy
P.S. The discussion on this thread can be found at [2] [3].

[1] http://download.java.net/jdk7/docs/platform/jvmti/jvmti.html#GetThreadState
[2] http://mail.openjdk.java.net/pipermail/core-libs-dev/2010-July/004567.html
[3] http://mail.openjdk.java.net/pipermail/core-libs-dev/2010-August/004721.html


	JDK 7 b120 (in ms)	With fix (in ms)	Speed up
main		46465	        22772			2.04
NEW		50676		29921			1.69
RUNNABLE	42202		14690			2.87
BLOCKED		72773		12296			5.92
WAITING		48811		13041			3.74
TIMED_WAITING	45737		12849			3.56
TERMINATED	40314		16376			2.46

Reply via email to