Java Garbage Collectors

Generational Collectors (分代收集器)

• GC algos optimised based on two hypotheses / observations:

• Most objects soon become unreachable - short lived.

• References from old objects to young objects only exist in small numbers

• The Oracle HotSpot JVM:

• Objects allocated in the Eden space of the Young Generation (or New Generation年轻代)

• Once the Eden space is full, a young collection or minor collection occurs

• Surviving objects live in the survivor space of the young generation

• When an object is “old enough”, it is promoted to the Old Generation (or tenured space年老代)

• When the old generation is “full enough”, a major collection occurs.

• Allocation is usually very fast

• Thread Local Allocation Buffers

• Bump of pointer

Java Garbage Collectors

• Available in the Oracle HotSpot JVM:

• Serial Collector

• Parallel (Throughput) Collector

• Concurrent Mark-Sweep Collector (CMS)

• G1

• Others:

• Oracle JRockit Real Time

• IBM Websphere Real Time

• Azul Zing

Serial Collector

• Single threaded for all collections (minor and major)

• All collections are Stop-The-World (STW)

• Collections use a mark-sweep-compact algorithm

• Suitable for single-threaded apps with low memory footprint (~100MB)

• Enabled using -XX:+UseSerialGC

Parallel Collector

• Similar to the serial collector, but uses multiple threads to speed things up

• Offers highest throughput of all the collectors

• Enabled using ‑XX:+UseParallelGC

Concurrent Mark Sweep (CMS) Collector

• Boasts shorter pauses than the serial or parallel collectors at the expense of application throughput

• Minor collections similar to serial collector

• Old generation is collected concurrently with the application

• Not compacting, so can result in old generation fragmentation

• Enabled using -XX:+UseConcMarkSweepGC

G1 Garbage Collector

• G1 = Garbage First

• Default GC in Java 9

• Aims for low pause times (<0.5s) and good throughput (90%) for large heaps (>6GB)

• Generational and concurrent

• Adaptive to meet pause time target

• Enabled with -XX:+UseG1GC

G1 Layout

• Unlike other collectors, G1 divides the heap into evenly sized regions

• Regions can be 1MB to 32MB in size, in power of two increments

• Regions can be dynamically assigned as:

• Eden

• Survivor

• Old

• Humongous

• G1 aims for 2048 regions based on minimum heap size

G1 Minor Collections

• Also known as Evacuation Pauses

• A STW event

• Subset of regions logically assigned as the young generation

• Minor collection triggered when the young generation is full

• Live objects “evacuated” to new regions to achieve compaction

• Objects moved to either old region or survivor region based on age

• Number of regions in young generation can be changed to meet the pause time target

G1 Concurrent Marking

• Triggered when the used heap reaches a configurable threshold of total heap

• Aim is to identify which old generation regions can be collected

• Multi-phased process, some of which is STW, some concurrent with application

• Concurrent phase can be stopped by a young collection

G1 Mixed Collections

• Occurs after the concurrent marking phase

• Old regions optionally added to the eden and survivor regions to be collected

• Old regions eligible for collection usually split over multiple collections

• The number of mixed collections is tunable via flags

• G1 reverts to minor (young) collections when mixed collections have finished

G1 Humongous Objects

• Humongous object defined as one greater than 50% the size of a region.

• Allocated directly to the old generation to avoid copying during young collections

• Region(s) marked as humongous

• Don’t want too many, and ideally they should be long lived.

G1 Evacuation Failures and Full GC

• Evacuation failure is when there is no free space to copy objects to

• Evacuation failures trigger a full GC - very expensive!

G1 Configuration

• There are many flags. Advice is to not tune G1 much unless you have to.

• Some flags are experimental, and require -XX:+UnlockExperimentalVMOptions to be set also

• Primary flag is to control the pause target:
  -XX:MaxGCPauseMillis=200

• -XX:InitiatingHeapOccupancyPercent=45
  Percentage of heap occupancy at which the marking phase is triggered

• Many more options described at http://www.oracle.com/technetwork/articles/java/g1gc-1984535.html

• Some options for GC logging:

• -XX:+PrintGCDetails - Enable detailed GC logging

• -XX:+PrintGCDateStamps - Print absolute date stamps at beginning of GC lines

• -XX:+PrintGCTimeStamps Print a timestamp reflecting the real time passed in seconds since JVM start

GC Logs

• New log file every time the JVM starts