10 minutes to understand CMS garbage collector in JVM

CMS It's based on markers - Clearing algorithm's , The collection is divided into 4 A step ：

1. Initial marker

2. Concurrent Tags

3. Re label

4. Concurrent elimination

Initial marker

The initial mark is just a mark GC Roots Objects that can be directly related to , So it's very fast . For example, below , This way GC Roots Using the virtual machine stack as an example . Two VMS stack tables create objects OBJ_A1 and OBJ_B1, They also have their own other quotes , At this stage , He just marks OBJ_A1 and OBJ_B1, Other references are unmarked , So even though this stage has STW, But the number of markers is small , Time soon , It doesn't affect .

Concurrent Tags

Concurrent marking is based on the object directly or indirectly referenced by the initial marked object , For example, the figure below is right OBJ_A2,OBJ_AN Mark concurrency . There's no such thing at this stage STW, So you can create objects , Add a new reference , It also causes some objects to lose references , For example, below ,OBJ_B1 It has become rubbish ,OBJ_C1 It's a new live object . This stage is due to tracking all the objects of the old age , So it's very time consuming .

Re label

In concurrent tags , We see living objects OBJ_C1 And garbage objects OBJ_B1 Wait is not marked , So this stage is to re label these objects . There are also STW, However, only objects with changes in concurrent tags are marked , These numbers are relatively small , So the speed is also very fast .

Concurrent elimination

This stage , It's just after the tagging , Cleaning up garbage objects , Just like concurrent tags , It's time consuming , Because there is no STW, So it has little effect on the operation of the program .CMS It uses the marking and clearing algorithm .

shortcoming

CMS Of 4 Stages , Initial marking and relabeling require STW, But the time is short , The impact is not big . Concurrent marking and concurrent cleanup do not require STW, Although it takes time , But concurrent execution , It doesn't matter much , look CMS Perfect. , But he also has some shortcomings .

CPU

CMS The number of recycle threads started by default is （CPU Number +3）/ 4, That is to say CPU stay 4 More than one time , The garbage collection thread shall not be less than 25% Of CPU resources , And with CPU An increase in quantity leads to a decrease . For example, the server is 2 nucleus 4G, Then we need to use (2+3)/4=1 Threads to handle concurrent tagging and concurrent cleanup , At this time, only 1 Threads handle other things .

Floating garbage

The production of floating garbage , Mainly in the concurrent cleanup phase . After re labeling ,CMS The garbage collector will know what needs to be cleaned up , In the concurrent cleanup phase , Clean up the re marked garbage objects , There's no such thing at this stage STW, So it's possible to create new objects .

For example, in the picture below OBJ_N, After you create , Stack frames are recycled , The quotation is gone , He can't be cleared at this stage , I can only wait for the next garbage collection , Marked and cleared .

If the objects in the old age of this stage exceed the remaining space , Will appear Concurrent Mode Failure Failure , The virtual opportunity is temporarily enabled Serial Old Collector for old age garbage collection .

It can be used XX:CMSInitiatingOccupancyFraction Set the percentage of old age space occupied CMS Recycling ,JDK1.6 After the default 92%.

space debris

stay 《JVM Garbage collection algorithm 》 It is mentioned in the article , Mark - The clean-up algorithm will generate space debris , If there is not enough contiguous memory space to hold objects that are about to enter the old age , This will trigger Full GC.

To avoid that ,CMS The collector provides a -XX:+UseCMSCompactAtFullCollection, Open by default , When Full GC After completion , He will STW, To organize the memory , Keep the living objects together , Make room for continuity .

If you have to reset the memory every time , The city STW, therefore CMS It also provides -XX:CMSFullGCsBeforeCompaction Parameters , The default is 0, Indicates how many times Full GC After finishing the memory .

Link to the original text ：https://mp.weixin.qq.com/s?__biz=Mzg4MjU0OTM1OA==&mid=2247499410&idx=1&sn=787037527afd5762e407626d010a7589