[
https://issues.apache.org/jira/browse/SOLR-6734?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16505714#comment-16505714
]
Shawn Heisey commented on SOLR-6734:
------------------------------------
Putting a bit of code in a comment so that I don't lose track of it.
{code:java}
package foo;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicInteger;
import java.util.concurrent.atomic.AtomicLong;
public class Main {
private static final AtomicInteger threadCounter = new AtomicInteger(0);
public static void main(String[] args) throws InterruptedException {
Thread gce = new GCExerciser();
/*
* To see the difference made by the GC exerciser, run the program once
as-is,
* and then again with the following start() call commented.
*/
gce.start();
Thread pd = new PauseDetector();
pd.start();
}
/**
* Background thread to allocate a byte array every 10 milliseconds. This
serves
* to create garbage, so that there will be GC pauses.
*/
public static class GCExerciser extends Thread {
@Override
public void run() {
Thread.currentThread().setName("GCExerciser" +
threadCounter.incrementAndGet());
final long maxHeap = Runtime.getRuntime().maxMemory();
final long thirtyOneGig = 31L * 1024 * 1024 * 1024;
final int size;
if (maxHeap > thirtyOneGig) {
size = Integer.MAX_VALUE - 1;
} else {
size = (int) (maxHeap / 16);
}
byte[] b = null;
while (true) {
try {
Thread.sleep(10);
} catch (InterruptedException e) {
//
}
b = new byte[size];
@SuppressWarnings("unused")
int i = b.length;
}
}
}
public static class PauseDetector extends Thread {
private static final AtomicLong accumulatedNanos = new AtomicLong(0);
private static final int sleepMillis = 500;
private static final long sleepNanos =
TimeUnit.NANOSECONDS.convert(sleepMillis, TimeUnit.MILLISECONDS);
private static final long fiveMinuteNanos = TimeUnit.NANOSECONDS.convert(5,
TimeUnit.MINUTES);
@Override
public void run() {
Thread.currentThread().setName("PauseDetector" +
threadCounter.incrementAndGet());
Thread.currentThread().setPriority(Thread.MIN_PRIORITY);
long fiveMinuteStart = System.nanoTime();
while (true) {
long sleepStart = System.nanoTime();
try {
Thread.sleep(sleepMillis);
} catch (InterruptedException e) {
//
}
long sleepStop = System.nanoTime();
long sleepElapsed = sleepStop - sleepStart;
/*
* If elapsed time is more than the half second sleep, add the
difference to the
* accumulation counter.
*/
if (sleepElapsed > sleepNanos) {
accumulatedNanos.addAndGet(sleepElapsed - sleepNanos);
}
/*
* If total elapsed time is five minutes or longer, output the
accumulated time
* and reset everything.
*/
if ((sleepStop - fiveMinuteStart) >= fiveMinuteNanos) {
System.out.printf("Accumulated pause estimate for five minutes: %d
ms\n",
TimeUnit.MILLISECONDS.convert(accumulatedNanos.get(),
TimeUnit.NANOSECONDS));
fiveMinuteStart = System.nanoTime();
accumulatedNanos.set(0);
}
}
}
}
}
{code}
The PauseDetector thread in that code approximates what the jHiccup tool from
Azul does.
https://www.azul.com/products/open-source-tools/jhiccup-performance-tool/
With a version of this thread in Solr and any other services, it would be very
easy to detect and log severe application pauses, which typically are caused by
GC.
> Standalone solr as *two* applications -- Solr and a controlling agent
> ---------------------------------------------------------------------
>
> Key: SOLR-6734
> URL: https://issues.apache.org/jira/browse/SOLR-6734
> Project: Solr
> Issue Type: Sub-task
> Reporter: Shawn Heisey
> Priority: Major
>
> In a message to the dev list outlining reasons to switch from a webapp to a
> standalone app, Mark Miller included the idea of making Solr into two
> applications, rather than just one. There would be Solr itself, and an agent
> to control Solr.
> http://mail-archives.apache.org/mod_mbox/lucene-dev/201305.mbox/%3C807476C6-E4C3-4E7E-9F67-2BECB63990DE%40gmail.com%3E
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
