Skip to content

Garbage Collection

The GC module registers with the notification emitter of the GarbageCollectorMXBean to provide some basic GC logging and metrics.

Getting Started

For using it internally at Netflix, see the Java Usage guide, otherwise keep reading this section.

Requirements

This library relies on the notification emitter added in 7u4, but there are known issues prior to 7u40. There is also a regression impacting Java 9 and higher, see #502 and JDK-8196325 for more information. For G1, it is recommended to be on the latest version available.

Dependencies

com.netflix.spectator:spectator-ext-gc:0.101.0

Start Reporting

Then in the initialization for the application:

import com.netflix.spectator.gc.GcLogger;
...
// Keep a single instance of the logger
GcLogger gc = new GcLogger();
gc.start(null);

Logging

After GC events, an DEBUG level log message will get reported using slf4j. This makes it easy to see GC events in the context of other log messages for the application. The logger name is com.netflix.spectator.gc.GcLogger and the message will look like:

${GC_TYPE}: ${COLLECTOR_NAME}, id=${N}, at=${START_TIME}, duration=${T}ms,
cause=[${CAUSE}], ${TOTAL_USAGE_BEFORE} => ${TOTAL_USAGE_AFTER} / ${MAX_SIZE}
(${PERCENT_USAGE_BEFORE} => ${PERCENT_USAGE_AFTER})

The id can be used to verify events were not skipped or correlate with other sources like detailed GC logs. See GC causes for more details on the possible causes.

Sample:

2014-08-31 02:02:24,724  DEBUG [com.netflix.spectator.gc.GcLogger] YOUNG: ParNew,
id=5281, at=Sun Aug 31 02:02:24 UTC 2014, duration=2ms, cause=[Allocation Failure],
0.4G => 0.3G / 1.8G (24.3% => 16.6%)

Metrics

jvm.gc.allocationRate

The allocation rate measures how fast the application is allocating memory. It is a counter that is incremented after a GC event by the amount youngGen.sizeBeforeGC.

Technically, right now it is:

youngGen.sizeBeforeGC - youngGen.sizeAfterGC

However, youngGen.sizeAfterGC should be 0 and thus the size of young gen before the GC is the amount allocated since the previous GC event.

Unit: bytes/second

Dimensions:

  • None.

jvm.gc.promotionRate

The promotion rate measures how fast data is being moved from young generation into the old generation.

It is a counter that is incremented after a GC event by the amount:

abs(oldGen.sizeAfterGC - oldGen.sizeBeforeGC)

Unit: bytes/second

Dimensions:

  • None.

jvm.gc.liveDataSize

The live data size is the size of the old generation after a major GC.

The image below shows how the live data size view compares to a metric showing the current size of the memory pool:

Live Data Size

Unit: bytes

Dimensions:

  • None.

jvm.gc.maxDataSize

Maximum size for the old generation. Primary use-case is for gaining perspective on the the live data size.

Unit: bytes

Dimensions:

  • None.

jvm.gc.pause

Pause time for a GC event. All of the values reported are stop the world pauses.

Unit: seconds

Dimensions:

  • action: action performed by the garbage collector (getGcAction). There is no guarantee, but the typical values seen are end_of_major_GC and end_of_minor_GC.
  • cause: cause that instigated GC (getGcCause). For an explanation of common causes see the GC causes page.

jvm.gc.concurrentPhaseTime

Time spent in concurrent phases of CMS pauses.

Unit: seconds

Dimensions:

  • action: action performed by the garbage collector (getGcAction). There is no guarantee, but the typical values seen are end_of_major_GC and end_of_minor_GC.
  • cause: cause that instigated GC (getGcCause). For an explanation of common causes see the GC causes page.

Alerting

This section assumes the data is available in Atlas, but users of other systems should be able to take the idea and make it work. For all of these alerts it is recommended to check them on instance. At Netflix that can be done by selecting the option in alert ui:

On Instance Alerting

Max Pause Time

Example to trigger an alert if the pause time exceeds 500 milliseconds:

name,jvm.gc.pause,:eq,
statistic,max,:eq,
:and,
:max,(,cause,),:by,
0.5,:gt,
$cause,:legend

Heap Pressure

Example to trigger an alert if the live data size is over 70% of the heap:

name,jvm.gc.liveDataSize,:eq,:max,
name,jvm.gc.maxDataSize,:eq,:max,
:div,100,:mul,
70,:gt,
percentUsed,:legend