ParallelGC issue: collector does only Full GC
Srinivas Ramakrishna
ysr1729 at gmail.com
Thu Oct 24 13:00:17 PDT 2013
Just looked at the publicly visible description of 7098155, but it doesn't
directly talk about the "freeze" in input telemetry
of promotion volume. Just about the need to resize at a full gc.
That latter to me is a somewhat orthogonal (but not unrelated) issue.
May be the non-public parts of the CR talk about the lack of new telemetry
which would correct the promotion estimate/prediction.
-- ramki
On Thu, Oct 24, 2013 at 12:28 PM, Srinivas Ramakrishna <ysr1729 at gmail.com>wrote:
> Yes, exactly: that's point (3) of my email from 5/9, a pointer to which I
> included earlier in this (slightly bifuracted) thread.
> Thanks for fixing it!
>
> I'd submit that the other two points, especially point (1), should also be
> done.
> -- ramki
>
>
> On Thu, Oct 24, 2013 at 12:01 PM, Jon Masamitsu <jon.masamitsu at oracle.com>wrote:
>
>>
>> On 10/24/13 5:00 AM, charlie hunt wrote:
>>
>> I did a little experimenting with this ... I think Jon's hypothesis is
>> right.
>>
>> I first reproduced the behavior as described by Andreas. Then, I set
>> -XX:PromotedPadding=1 in the case where NewSize/MaxNewSize is at 1800m.
>> That eliminated the issue Andreas observed at 1800m. But, as I suspected
>> the threshold at which the change in behavior merely changed at a higher
>> sizing of young gen. It now occurs at about 2300m, up from 1800m.
>>
>> So, it does look like there is an issue with the prediction model since
>> PromotedPadding can influence the prediction model.
>>
>> The prediction model code does not look trivial, as I'm sure Jon knows.
>> ;)
>>
>>
>> Part of the problem is that some of the inputs to the prediction model
>> are not updated at a Full GC. So depending on how much space
>> gets freed up during the Full GC, the prediction model can get
>> frozen into the wrong decision. That's what we're trying to fix
>> with 7098155.
>>
>> Jon
>>
>>
>>
>> hths,
>>
>> charlie ...
>>
>>
>>
>>
>> On Wed, Oct 23, 2013 at 3:03 AM, Andreas Müller <
>> Andreas.Mueller at mgm-tp.com> wrote:
>>
>>> Hi Jon,
>>>
>>> thanks for the hint to Java 8.
>>> I have verified with jdk1.8.0-ea-b112 (from October 17): behavior
>>> remains as described
>>>
>>> Best regards
>>> Andreas
>>>
>>> ----------------------------------------------------------------------
>>>
>>> Date: Mon, 21 Oct 2013 15:29:10 -0700
>>> From: Jon Masamitsu <jon.masamitsu at oracle.com>
>>> Subject: Re: ParallelGC issue: collector does only Full GC by default
>>> and above NewSize=1800m
>>> To: hotspot-gc-use at openjdk.java.net
>>> Message-ID: <5265AAB6.1050700 at oracle.com>
>>> Content-Type: text/plain; charset="iso-8859-1"
>>>
>>> Andreas,
>>>
>>> There was a bug fixed in jdk8 that had similar symptoms. If you can try
>>> a jdk8 build that might tell us something.
>>>
>>> If jdk8 doesn't help it's likely that the prediction model thinks that
>>> there is not enough
>>> free space in the old gen to support a young collection. We've been
>>> working on 7098155 to
>>> fix that.
>>>
>>> Jon
>>>
>>>
>>> On 10/21/2013 10:09 AM, Andreas M?ller wrote:
>>> > Hi all,
>>> >
>>> > while experimenting a bit with different Garbage Collectors and
>>> > applying them to my homegrown micro benchmarks I stumbled into the
>>> following problem:
>>> > I run the below sample with the following command line (using Java
>>> 1.7.0_40 on Windows and probably others):
>>> > java -Xms6g -Xmx6g -XX:+UseParallelGC -
>>> > de.am.gc.benchmarks.MixedRandomList 100 8 12500000
>>> >
>>> > The Default and proven ParallelGC collector does mostly Full GCs and
>>> shows only poor out-of-the-box performance, more than a factor 10 lower
>>> than the ParNew collector.
>>> > More tests adding the -XX:NewSize=<xyz>m and -XX:MaxNewSize=<xyz>m
>>> reveal that the problem occurs as soon as the NewSize rises beyond 1800m
>>> which it obviously does by default.
>>> > Below that threshold ParallelGC performance is similar to ParNewGC (in
>>> the range of 7500 MB/s on my i7-2500MHz notebook), but at NewSize=2000m is
>>> as low as 600 MB/s.
>>> >
>>> > Any ideas why this might happen?
>>> >
>>> > Note that the sample is constructed such that the live heap is always
>>> around 3GB. If any I would expect a problem only at around NewSize=3GB,
>>> when Old Gen shrinks to less than the live heap size. As a matter of fact,
>>> ParNewGC can do >7000 MB/s from NewSize=400m to NewSize=3500m with little
>>> variation around a maximum of 7600 MB/s at NewSize=2000m.
>>> >
>>> > I also provide source, gc.log and a plot of the NewSize dependency to
>>> anyone interested in that problem.
>>> >
>>> > Regards
>>> > Andreas
>>> >
>>> > -------------------------------------------------------MixedRandomList
>>> > .java-----------------------------------------------------------------
>>> > -------------------------------------------------------
>>> > package de.am.gc.benchmarks;
>>> >
>>> > import java.util.ArrayList;
>>> > import java.util.List;
>>> >
>>> > /**
>>> > * GC benchmark producing a mix of lifetime=0 and lifetime>0 objects
>>> which are kept in randomly updated lists.
>>> > *
>>> > * @author Andreas Mueller
>>> > */
>>> > public class MixedRandomList {
>>> > private static final int DEFAULT_NUMBEROFTHREADS=1;
>>> > // object size in bytes
>>> > private static final int DEFAULT_OBJECTSIZE=100;
>>> >
>>> > private static int numberOfThreads=DEFAULT_NUMBEROFTHREADS;
>>> > private static int objectSize=DEFAULT_OBJECTSIZE;
>>> > // number of objects to fill half of the available memory with
>>> (permanent) live objects
>>> > private static long numLive =
>>> > (Runtime.getRuntime().maxMemory()/objectSize/5);
>>> >
>>> > /**
>>> > * @param args the command line arguments
>>> > */
>>> > public static void main(String[] args) {
>>> > if( args.length>0 ) {
>>> > // first, optional argument is the size of the objects
>>> > objectSize = Integer.parseInt(args[0]);
>>> > // second, optional argument is the number of live objects
>>> > if( args.length>1 ) {
>>> > numberOfThreads = Integer.parseInt(args[1]);
>>> > // third, optional argument is the number of live
>>> objects
>>> > if( args.length>2 ) {
>>> > numLive = Long.parseLong(args[2]);
>>> > }
>>> > }
>>> > }
>>> > for( int i=0; i<numberOfThreads; i++ ) {
>>> > // run several GarbageProducer threads, each with its own
>>> mix of lifetime=0 and higher lifetime objects
>>> > new Thread(new
>>> GarbageProducer((int)Math.pow(50.0,(double)(i+1)),
>>> numLive/numberOfThreads)).start();
>>> > }
>>> > try {
>>> > Thread.sleep(1200000);
>>> > } catch( InterruptedException iexc) {
>>> > iexc.printStackTrace();
>>> > }
>>> > System.exit(0);
>>> > }
>>> >
>>> > private static char[] getCharArray(int length) {
>>> > char[] retVal = new char[length];
>>> > for(int i=0; i<length; i++ ) {
>>> > retVal[i] = 'a';
>>> > }
>>> > return retVal;
>>> > }
>>> >
>>> > public static class GarbageProducer implements Runnable {
>>> >
>>> > // the fraction of newly created objects that do not become
>>> garbage immediately but are stored in the liveList
>>> > int fractionLive;
>>> > // the size of the liveList
>>> > long myNumLive;
>>> >
>>> > /**
>>> > * Each GarbageProducer creates objects that become garbage
>>> immediately (lifetime=0) and
>>> > * objects that become garbage only after a lifetime>0 which
>>> is distributed about an average lifetime.
>>> > * This average lifetime is a function of fractionLive and
>>> numLive
>>> > *
>>> > * @param fractionLive
>>> > * @param numLive
>>> > */
>>> > public GarbageProducer(int fractionLive, long numLive) {
>>> > this.fractionLive = fractionLive;
>>> > this.myNumLive = numLive;
>>> > }
>>> >
>>> > @Override
>>> > public void run() {
>>> > int osize = objectSize;
>>> > char[] chars = getCharArray(objectSize);
>>> > List<String> liveList = new
>>> ArrayList<String>((int)myNumLive);
>>> > // initially, the lifeList is filled
>>> > for(int i=0; i<myNumLive; i++) {
>>> > liveList.add(new String(chars));
>>> > }
>>> > while(true) {
>>> > // create the majority of objects as garbage
>>> > for(int i=0; i<fractionLive; i++) {
>>> > String garbageObject = new String(chars);
>>> > }
>>> > // keep the fraction of objects live by placing them
>>> in the list (at a random index)
>>> > int index = (int)(Math.random()*myNumLive);
>>> > liveList.set(index, new String(chars));
>>> > }
>>> > }
>>> > }
>>> > }
>>> > ----------------------------------------------------------------------
>>> > ----------------------------------------------------------------------
>>> > ---------------------------------------------------------------------
>>> >
>>> > Andreas M?ller
>>> >
>>> > mgm technology partners GmbH
>>> > Frankfurter Ring 105a
>>> > 80807 M?nchen
>>> > Tel. +49 (89) 35 86 80-633
>>> > Fax +49 (89) 35 86 80-288
>>> > E-Mail Andreas.Mueller at mgm-tp.com<mailto:Andreas.Mueller at mgm-tp.com>
>>> > Innovation Implemented.
>>> > Sitz der Gesellschaft: M?nchen
>>> > Gesch?ftsf?hrer: Hamarz Mehmanesh
>>> > Handelsregister: AG M?nchen HRB 105068
>>> >
>>> >
>>> >
>>> >
>>> > _______________________________________________
>>> > hotspot-gc-use mailing list
>>> > hotspot-gc-use at openjdk.java.net
>>> > http://mail.openjdk.java.net/mailman/listinfo/hotspot-gc-use
>>>
>>> -------------- next part --------------
>>> An HTML attachment was scrubbed...
>>> URL:
>>> http://mail.openjdk.java.net/pipermail/hotspot-gc-use/attachments/20131021/62d044c7/attachment-0001.html
>>>
>>> ------------------------------
>>>
>>> _______________________________________________
>>> hotspot-gc-use mailing list
>>> hotspot-gc-use at openjdk.java.net
>>> http://mail.openjdk.java.net/mailman/listinfo/hotspot-gc-use
>>>
>>>
>>> End of hotspot-gc-use Digest, Vol 68, Issue 5
>>> *********************************************
>>> _______________________________________________
>>> hotspot-gc-use mailing list
>>> hotspot-gc-use at openjdk.java.net
>>> http://mail.openjdk.java.net/mailman/listinfo/hotspot-gc-use
>>>
>>
>>
>>
>> _______________________________________________
>> hotspot-gc-use mailing list
>> hotspot-gc-use at openjdk.java.net
>> http://mail.openjdk.java.net/mailman/listinfo/hotspot-gc-use
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-gc-use/attachments/20131024/177c3faa/attachment-0001.html
More information about the hotspot-gc-use
mailing list