G1原理—7.G1的GC日志分析解读
大纲
1.TLAB的GC日志解读
2.YGC的GC日志解读
3.模拟YGC(单次GC及多次GC的不同场景)
4.打开实验选项查看YGC的详情日志信息
5.Mixed GC日志信息之初始标记过程
6.Mixed GC日志信息之混合回收过程
7.Mixed GC日志信息之Region的详细信息和标记过程的详细信息
8.FGC的日志信息及日志详情
1.TLAB的GC日志解读
(1)一套基本的参数设置
(2)针对这个参数设置写一套简单的Demo代码
(3)TLAB相关的运行日志
(1)一套基本的参数设置
首先参数上要有一套基本的设置:
-XX:InitialHeapSize=128M -XX:MaxHeapSize=128M -XX:+UseG1GC
-XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+PrintTLAB
-XX:+UnlockExperimentalVMOptions -XX:G1LogLevel=finest
-XX:MaxGCPauseMillis=20 -Xloggc:gc.log
堆内存的大小设置为128M是因为:G1本身的Region都要最小1M,总共就没有几个Region。如果设置成10MB这种值,不好做实验。
-XX:+UnlockExperimentalVMOptions这个参数是一个实验选项,打开以后才可以使用-XX:G1LogLevel这个参数,-XX:G1LogLevel=finest表示设置日志的输出级别为最详细级别。
(2)针对这个参数设置写一套简单的Demo代码
public class TLABDemo {
private static final ArrayList<String> strs = new ArrayList<String>();
public static void main(String[] args) throws Exception {
for (int n = 0 ;; n++) {
//外层循环100次,每次内层循环创建10个字符串
for (int i = 0; i<100; i++) {
//内层循环10次
for (int j = 0; j < 10; j++) {
strs.add(new String("NO." + j + "Str" ));
}
}
//无限循环,每次循环间隔0.1s
System.out.println("第" + n + "次循环");
Thread.sleep(100);
}
}
}
这段代码的意思是:无限循环去创建字符串,放到一个list去。每次循环会执行1000次创建字符串,并把字符串加入到list的操作。通过这段代码,看看TLAB相关的GC日志信息。
(3)TLAB相关的运行日志
//这一部分是我们设置的一些参数信息
CommandLine flags: -XX:G1LogLevel=finest -XX:InitialHeapSize=134217728 -XX:MaxGCPauseMillis=20 -XX:MaxHeapSize=134217728 -XX:+PrintGC -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+PrintTLAB -XX:+UnlockExperimentalVMOptions -XX:+UseCompressedClassPointers -XX:+UseCompressedOops -XX:+UseG1GC -XX:-UseLargePagesIndividualAllocation
2.236: [GC pause (G1 Evacuation Pause) (young)
TLAB: gc thread: 0x00000000181fd800 [id: 8648] desired_size: 122KB slow allocs: 0 refill waste: 1960B alloc: 0.99996 6144KB refills: 1 waste 1.4% gc: 1704B slow: 0B fast: 0B
TLAB: gc thread: 0x00000000181fa000 [id: 3280] desired_size: 122KB slow allocs: 2 refill waste: 1960B alloc: 0.99996 6144KB refills: 3 waste 25.0% gc: 91168B slow: 3104B fast: 0B
TLAB: gc thread: 0x00000000181bb000 [id: 10280] desired_size: 122KB slow allocs: 2 refill waste: 1960B alloc: 0.99996 6144KB refills: 7 waste 2.0% gc: 13992B slow: 3624B fast: 0B
TLAB: gc thread: 0x0000000003683800 [id: 12164] desired_size: 122KB slow allocs: 6 refill waste: 1960B alloc: 0.99996 6144KB refills: 41 waste 0.1% gc: 0B slow: 7584B fast: 0B
TLAB totals: thrds: 4 refills: 52 max: 41 slow allocs: 10 max 6 waste: 1.9% gc: 106864B max: 91168B slow: 14312B max: 7584B fast: 0B max: 0B, 0.0025449 secs]
//这一部分是我们设置的一些参数信息
CommandLine flags: -XX:G1LogLevel=finest -XX:InitialHeapSize=134217728 -XX:MaxGCPauseMillis=20 -XX:MaxHeapSize=134217728 -XX:+PrintGC -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+PrintTLAB -XX:+UnlockExperimentalVMOptions -XX:+UseCompressedClassPointers -XX:+UseCompressedOops -XX:+UseG1GC -XX:-UseLargePagesIndividualAllocation
//下面的GC pause,就是我们的停顿时间,后面跟的young指的是新生代
2.236: [GC pause (G1 Evacuation Pause) (young)
//TLAB后跟的这一大段,其中有几个关键点需要我们注意
TLAB: gc thread: 0x00000000181fd800 [id: 8648] desired_size: 122KB slow allocs: 0 refill waste: 1960B alloc: 0.99996 6144KB refills: 1 waste 1.4% gc: 1704B slow: 0B fast: 0B
gc thread:表示是哪个线程的TLAB。
desired_size:122KB指的是期望分配的TLAB的大小。这个值,其实就是我们前面讲解TLAB计算公式的时候,计算出来的值。TLABSize = Eden * 2 * 1% / 线程个数。
slow allocs:0,指的是慢速分配的次数。0说明每次都使用了TLAB快速分配,没有直接使用堆内存去分配。注意慢速TLAB分配也算是快速分配。
refill waste:1960B,代表可浪费的内存,也是重新申请一个TLAB的阈值。
alloc: 0.99996,表示当前这个线程在一个分区中分配对象占用的比例,Region使用的比例。
refills: 1,代表出现多少次废弃TLAB,重新申请一个TLAB的次数。refills次数为1表示只做了一次TLAB申请。每次废弃TLAB时,都会填充一个dummy对象。注意:refills的次数就是填充TLAB的次数,可理解为申请新的TLAB的次数。
waste 1.4% gc: 1704B slow: 0B fast: 0B,waste代表的是浪费的空间,这个浪费的空间分为三个层面:
一.gc表示:GC时还没有使用的TLAB的空间,现在正处于GC的状态中,TLAB还剩下多少空间没使用。
二.slow表示:申请新的TLAB时,旧的TLAB浪费的空间。这里就是0B,因为只产生了1次。
三.fast表示:在出现需要调整TLAB的大小时,即refill_waste不合理时,旧的TLAB浪费的空间,也就是dummy对象造成的浪费。
2.236: [GC pause (G1 Evacuation Pause) (young)TLAB: gc thread: 0x00000000181fd800 [id: 8648] desired_size: 122KB slow allocs: 0 refill waste: 1960B alloc: 0.99996 6144KB refills: 1 waste 1.4% gc: 1704B slow: 0B fast: 0B
观察这个TLAB状态,refill次数还是非常有用的。refill次数可以帮助我们调整TLAB相关参数的设置,有助于优化性能。
比如,经过压测通过GC日志发现:当TLAB的大小设置成 1/4 RegionSize时会出现更少的TLAB动态调整情况,并且这个值能够保证大多数对象能够直接走TLAB分配。那么就可手动设置TLAB的大小,不使用G1自动推断的值,避免动态调整。因为自动调整是需要代价的,每次自动调整都会有一定的性能损耗。
2.YGC的GC日志解读
(1)参数设置
(2)沿用简单的Demo代码
(3)代码及YGC时机分析
(4)YGC相关的运行日志
(1)参数设置
对于YGC来说,我们只是要模拟YGC的整个过程。并且要打印出YGC整个过程的一些GC细节,所以才设置这么一套参数:
-Xmx256M -XX:+UseG1GC
-XX:+PrintGCTimeStamps -XX:+PrintGCDetails
-XX:MaxGCPauseMillis=20 -Xloggc:gc.log
(2)沿用简单的Demo代码
public class TLABDemo {
private static final ArrayList<String> strs = new ArrayList<String>();
public static void main(String[] args) throws Exception {
for (int n = 0 ;; n++) {
//外层循环100次,每次内层循环创建10个字符串
for (int i = 0; i<100; i++) {
//内层循环10次
for (int j = 0; j < 10; j++) {
strs.add(new String("NO." + j + "Str" ));
}
}
//无限循环,每次循环间隔0.1s
System.out.println("第" + n + "次循环");
Thread.sleep(100);
}
}
}
(3)代码及YGC时机分析
首先,运行这个代码时,没有设置新生代占用多少空间就会发生YGC,所以具体什么时候会发生YGC暂时不确定。
不过可以简单的思考YGC的发生时机,尤其是第一次。第一次发生YGC时,应该是256M * 5%的新生代被使用的时候。大概是这个值,当然,实际可能会比这个值略小一点。因为新生代第一次GC时Survivor区是没有存活对象的,并且堆内存会有一部分被其他的一些描述信息占用,所以256M堆内存不会全部都用来分配对象。程序运行一段时间后:大量字符串对象会填满Eden区,此时就会触发一次YGC。
(4)YGC相关的运行日志
可以看到,G1的GC日志看起来非常复杂,中间有非常复杂的处理步骤。比如,roots扫描,RS扫描等操作。每一步的操作都会有各自的耗时时长。接下来一步一步分析这个日志。
//这一部分是我们设置的一些参数信息
CommandLine flags: -XX:InitialHeapSize=266064256 -XX:MaxGCPauseMillis=20 -XX:MaxHeapSize=268435456 -XX:+PrintGC -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+UseCompressedClassPointers -XX:+UseCompressedOops -XX:+UseG1GC -XX:-UseLargePagesIndividualAllocation
//从这里开始,出现了第一次GC,大概是在程序启动6.193s之后,停顿时间是0.0048641秒,也就是4.8ms左右;
6.193: [GC pause (G1 Evacuation Pause) (young), 0.0048641 secs]
//这里很明显看到GC线程并行工作的时间是2.9ms,并且使用了8个GCWorkers线程;
//因为此时电脑是8核的,不同的电脑,会有不同的结果,因为没有设置GC的GCParallelThreads线程数量的值;
[Parallel Time: 2.9 ms, GC Workers: 8]
//这些是线程启动,GC Roots扫描,RSet扫描,对象拷贝等操作的一些耗时情况;
[GC Worker Start (ms): Min: 6192.9, Avg: 6193.1, Max: 6193.6, Diff: 0.7]
[Ext Root Scanning (ms): Min: 0.0, Avg: 0.2, Max: 0.4, Diff: 0.4, Sum: 1.7]
[Update RS (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Processed Buffers: Min: 0, Avg: 0.0, Max: 0, Diff: 0, Sum: 0]
[Scan RS (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Code Root Scanning (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.1]
[Object Copy (ms): Min: 1.9, Avg: 2.2, Max: 2.3, Diff: 0.4, Sum: 17.8]
[Termination (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.1]
[Termination Attempts: Min: 16, Avg: 20.3, Max: 24, Diff: 8, Sum: 162]
[GC Worker Other (ms): Min: 0.0, Avg: 0.0, Max: 0.1, Diff: 0.0, Sum: 0.3]
[GC Worker Total (ms): Min: 1.9, Avg: 2.5, Max: 2.6, Diff: 0.7, Sum: 19.9]
[GC Worker End (ms): Min: 6195.5, Avg: 6195.5, Max: 6195.6, Diff: 0.0]
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.9 ms]
[Other: 1.0 ms]
[Choose CSet: 0.0 ms]
[Ref Proc: 0.5 ms]
[Ref Enq: 0.0 ms]
//卡表更新
[Redirty Cards: 0.5 ms]
//大对象相关的操作时间
[Humongous Register: 0.0 ms]
[Humongous Reclaim: 0.0 ms]
[Free CSet: 0.0 ms]
//在第一次GC之后,Eden区从使用12M,到使用0M,从总大小12M,到总大小10M,发生了动态变化;
//Survivor区从0变为了2048K,也就是本次GC之后有2048K的对象存活下来;
//堆内存从12M的使用量,降为了4424K的使用量,说明本次回收的效果还是不错的,基本上回收了百分之九十以上的垃圾;
[Eden: 12.0M(12.0M)->0.0B(10.0M) Survivors: 0.0B->2048.0K Heap: 12.0M(254.0M)->4424.0K(254.0M)]
[Times: user=0.00 sys=0.00, real=0.01 secs]
//第二次GC发生的时机,对于第二次GC,着重关注一下Eden Survivor 的大小,看看它是否有变化;
17.261: [GC pause (G1 Evacuation Pause) (young), 0.0087213 secs]
[Parallel Time: 7.7 ms, GC Workers: 8]
[GC Worker Start (ms): Min: 17261.5, Avg: 17261.5, Max: 17261.6, Diff: 0.2]
[Ext Root Scanning (ms): Min: 0.0, Avg: 0.1, Max: 0.3, Diff: 0.3, Sum: 1.0]
[Update RS (ms): Min: 0.0, Avg: 1.0, Max: 2.9, Diff: 2.9, Sum: 7.8]
[Processed Buffers: Min: 0, Avg: 1.3, Max: 3, Diff: 3, Sum: 10]
[Scan RS (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.1]
[Code Root Scanning (ms): Min: 0.0, Avg: 0.0, Max: 0.1, Diff: 0.1, Sum: 0.1]
[Object Copy (ms): Min: 4.6, Avg: 6.5, Max: 7.5, Diff: 2.9, Sum: 51.8]
[Termination (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.3]
[Termination Attempts: Min: 1, Avg: 2.6, Max: 5, Diff: 4, Sum: 21]
[GC Worker Other (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[GC Worker Total (ms): Min: 7.5, Avg: 7.6, Max: 7.7, Diff: 0.1, Sum: 61.1]
[GC Worker End (ms): Min: 17269.1, Avg: 17269.1, Max: 17269.1, Diff: 0.0]
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.3 ms]
[Other: 0.7 ms]
[Choose CSet: 0.0 ms]
[Ref Proc: 0.5 ms]
[Ref Enq: 0.0 ms]
[Redirty Cards: 0.1 ms]
[Humongous Register: 0.0 ms]
[Humongous Reclaim: 0.0 ms]
[Free CSet: 0.0 ms]
//从这里并没有发现新生代的变化,说明此时还没有进行新生代的调整;
//但是有一个现象,新生代使用从10MB降为了2048K,而堆内存,从14.9降为了12.9;
//也就是说,本次只回收了2M左右的对象,那么肯定是有一部分对象进入了老年代分区的;
[Eden: 10.0M(10.0M)->0.0B(10.0M) Survivors: 2048.0K->2048.0K Heap: 14.9M(254.0M)->12.9M(254.0M)]
[Times: user=0.09 sys=0.02, real=0.01 secs]
38.020: [GC pause (G1 Evacuation Pause) (young), 0.0101857 secs]
[Parallel Time: 8.9 ms, GC Workers: 8]
[GC Worker Start (ms): Min: 38019.9, Avg: 38020.0, Max: 38020.1, Diff: 0.1]
[Ext Root Scanning (ms): Min: 0.1, Avg: 0.3, Max: 1.3, Diff: 1.2, Sum: 2.2]
[Update RS (ms): Min: 0.0, Avg: 1.4, Max: 1.7, Diff: 1.7, Sum: 10.8]
[Processed Buffers: Min: 0, Avg: 1.9, Max: 3, Diff: 3, Sum: 15]
[Scan RS (ms): Min: 0.0, Avg: 0.1, Max: 0.5, Diff: 0.5, Sum: 0.7]
[Code Root Scanning (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Object Copy (ms): Min: 6.9, Avg: 7.0, Max: 7.1, Diff: 0.2, Sum: 56.4]
[Termination (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Termination Attempts: Min: 1, Avg: 28.9, Max: 43, Diff: 42, Sum: 231]
[GC Worker Other (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.1]
[GC Worker Total (ms): Min: 8.7, Avg: 8.8, Max: 8.8, Diff: 0.1, Sum: 70.3]
[GC Worker End (ms): Min: 38028.7, Avg: 38028.7, Max: 38028.7, Diff: 0.0]
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.3 ms]
[Other: 1.0 ms]
[Choose CSet: 0.0 ms]
[Ref Proc: 0.1 ms]
[Ref Enq: 0.0 ms]
[Redirty Cards: 0.7 ms]
[Humongous Register: 0.0 ms]
[Humongous Reclaim: 0.0 ms]
[Free CSet: 0.0 ms]
[Eden: 10.0M(10.0M)->0.0B(10.0M) Survivors: 2048.0K->2048.0K Heap: 25.2M(254.0M)->27.9M(254.0M)]
[Times: user=0.00 sys=0.00, real=0.01 secs]
58.728: [GC pause (G1 Evacuation Pause) (young), 0.0103349 secs]
[Parallel Time: 8.3 ms, GC Workers: 8]
[GC Worker Start (ms): Min: 58727.8, Avg: 58727.9, Max: 58728.0, Diff: 0.2]
[Ext Root Scanning (ms): Min: 0.0, Avg: 0.1, Max: 0.2, Diff: 0.2, Sum: 0.8]
[Update RS (ms): Min: 0.7, Avg: 1.0, Max: 1.5, Diff: 0.8, Sum: 8.3]
[Processed Buffers: Min: 1, Avg: 1.5, Max: 2, Diff: 1, Sum: 12]
[Scan RS (ms): Min: 0.0, Avg: 0.1, Max: 0.2, Diff: 0.2, Sum: 0.7]
[Code Root Scanning (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Object Copy (ms): Min: 6.3, Avg: 6.7, Max: 6.9, Diff: 0.6, Sum: 53.7]
[Termination (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Termination Attempts: Min: 1, Avg: 11.6, Max: 19, Diff: 18, Sum: 93]
[GC Worker Other (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.1]
[GC Worker Total (ms): Min: 7.8, Avg: 8.0, Max: 8.0, Diff: 0.2, Sum: 63.6]
[GC Worker End (ms): Min: 58735.8, Avg: 58735.8, Max: 58735.8, Diff: 0.0]
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.7 ms]
[Other: 1.2 ms]
[Choose CSet: 0.0 ms]
[Ref Proc: 0.5 ms]
[Ref Enq: 0.0 ms]
[Redirty Cards: 0.7 ms]
[Humongous Register: 0.0 ms]
[Humongous Reclaim: 0.0 ms]
[Free CSet: 0.0 ms]
[Eden: 10.0M(10.0M)->0.0B(10.0M) Survivors: 2048.0K->2048.0K Heap: 40.0M(254.0M)->42.0M(254.0M)]
[Times: user=0.00 sys=0.00, real=0.01 secs]
从多次GC的结果来看,新生代GC的情况相对比较稳定,每次的时长都在1ms左右(除了第一次时间比较长),并且由于我们的代码是没有对象置为null的操作的,所以基本上所有的对象都能存活下来晋升到老年代。
3.模拟YGC(单次GC及多次GC的不同场景)
(1)参数设置
(2)模拟代码
(3)代码运行相关的GC日志
(4)接下来把代码改造一下
(5)观察GC日志
(1)参数设置
还是这么一套参数结构,主要是为了和前面做一些对比,看看模拟出来的GC场景是否符合前面对日志的解读。
-Xmx256M -XX:+UseG1GC
-XX:+PrintGCTimeStamps -XX:+PrintGCDetails
-XX:MaxGCPauseMillis=20 -Xloggc:gc.log
(2)模拟代码
public class YGCDemo1 {
//5% * 256 = 12M左右是第一次发生YGC的时机
//因为堆内存256M / 2048小于1,根据G1的处理,会把一个Region定为1M
public static void main(String[] args) throws Exception {
//设定一个对象为256K的原因,如果太大,会直接走大对象分配无法模拟YGC
byte[] data = new byte[1024 * 256];
for (int i = 0; i< 36; i++) {
data = new byte[1024 * 256];
}
data = new byte[1024 * 256];
}
}
下面来解析一下代码:
一.首先第一段创建了一个data字节数组,大小是1024 * 256 = 256K
因为对象如果超过了Region大小一半,是一定会被当作大对象单独存储的,所以需要把这个对象的大小设置得合理一点。
而对于设置的这一套参数,堆内存是256M。通过计算不难发现,Region的大小是1M。因为256 / 2048很显然小于1M,然后会自动调整,调整成1M大小。代码中创建的对象设置成256K,就能正常分配到新生代Region中。
二.然后使用了一个循环,在循环中不断去创建新的数组
并且循环的次数设置在了36,为什么设置成36呢?原因是:G1启动时会初始化新生代的大小。新生代的大小占总空间的5%,也就是256M * 5% 约等于12M。也就是说,当应用程序用掉12个Region左右时,就会触发第一次YGC。
三.在触发了YGC之后
此时data这个变量还在,所以还是会有一个256K的字节数组存活,加上一些堆内存中存在的一些未知对象也可能会存活。那么在第一次回收后,应该会有不到1M的对象存活下来进入Survivor区域。
(3)代码运行相关的GC日志
//这些是我们设置的启动参数,其中需要注意的是,最大停顿时间,和堆内存的大小;
CommandLine flags: -XX:InitialHeapSize=266064256 -XX:MaxGCPauseMillis=20 -XX:MaxHeapSize=268435456 -XX:+PrintGC -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+UseCompressedClassPointers -XX:+UseCompressedOops -XX:+UseG1GC -XX:-UseLargePagesIndividualAllocation
//接下来看这里的内容,GC线程还是8个,本次GC造成的停顿时间是0.00178,很明显是在20ms内的;
0.155: [GC pause (G1 Evacuation Pause) (young), 0.0017840 secs] [Parallel Time: 1.1 ms, GC Workers: 8]
//这一段日志,可以看出来,在GC前,Eden区使用了12MB的内存,也就是12个Region;
//在本次GC之后,Survivor从GC开始前的0使用量,到了2048K,也就是Survivor使用了2个Region
//注意,这里不是说S区就有这么多的对象,而是S区总共使用了2个Region;
//堆内存的使用,从12M最后变为了1120K,大概1M左右
[Eden: 12.0M(12.0M)->0.0B(10.0M) Survivors: 0.0B->2048.0K Heap: 12.0M(254.0M)->1120.0K(254.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//堆内存的最终使用状态是这样的,总共256MB,使用了3MB左右,region size是1MB,5个region给了新生代,两个region给了survivor
Heap
garbage-first heap total 260096K, used 3168K [0x00000000f0000000, 0x00000000f01007f0, 0x0000000100000000)
region size 1024K, 5 young (5120K), 2 survivors (2048K)
Metaspace used 3451K, capacity 4496K, committed 4864K, reserved 1056768K
class space used 376K, capacity 388K, committed 512K, reserved 1048576K
(4)接下来把代码改造一下
public class YGCDemo1 {
private static final ArrayList<String> strs = new ArrayList<String>();
//5% * 256 = 12M左右
//因为256M / 2048小于1 ,根据G1的处理,会把一个Region定为1M;
public static void main(String[] args) throws Exception {
//设定一个对象为256K,如果太大,会直接走大对象分配无法模拟YGC
byte[] data = new byte[1024 * 256];
for (int i = 0; i< 36; i++) {
data = new byte[1024 * 256];
}
//创建n个对象,加入到一个list中,不回收这些数组;
ArrayList list = new ArrayList();
for (int i = 0; i<= 36; i++) {
byte[] data2 = new byte[1024 * 256];
//一部分字节数组加入到list中
if (i % 3 == 0) {
list.add(data2);
}
}
}
}
这段代码可以触发2次YGC,因为代码中会保留一些存活对象,所以第二次GC时,Region的使用情况会发生一些变化。
//创建n个对象,加入到一个list中,不回收这些数组;
ArrayList list = new ArrayList();
for (int i = 0; i<= 36; i++) {
byte[] data2 = new byte[1024 * 256];
//一部分字节数组加入到list中
if (i % 3 == 0) {
list.add(data2);
}
}
上面就是代码的变化位置,我们加了一个循环,并且每3个数组会有一个数组加入到list中不会被回收。
(5)观察GC日志
CommandLine flags: -XX:InitialHeapSize=266064256 -XX:MaxGCPauseMillis=20 -XX:MaxHeapSize=268435456 -XX:+PrintGC -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+UseCompressedClassPointers -XX:+UseCompressedOops -XX:+UseG1GC -XX:-UseLargePagesIndividualAllocation
0.150: [GC pause (G1 Evacuation Pause) (young), 0.0018467 secs]
[Parallel Time: 1.1 ms, GC Workers: 8]
[GC Worker Start (ms): Min: 149.9, Avg: 150.1, Max: 150.9, Diff: 0.9]
[Ext Root Scanning (ms): Min: 0.0, Avg: 0.2, Max: 0.4, Diff: 0.4, Sum: 1.7]
[Update RS (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Processed Buffers: Min: 0, Avg: 0.0, Max: 0, Diff: 0, Sum: 0]
[Scan RS (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Code Root Scanning (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Object Copy (ms): Min: 0.0, Avg: 0.5, Max: 0.6, Diff: 0.6, Sum: 3.9]
[Termination (ms): Min: 0.0, Avg: 0.1, Max: 0.1, Diff: 0.1, Sum: 0.5]
[Termination Attempts: Min: 1, Avg: 7.4, Max: 12, Diff: 11, Sum: 59]
[GC Worker Other (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.3]
[GC Worker Total (ms): Min: 0.0, Avg: 0.8, Max: 1.0, Diff: 0.9, Sum: 6.4]
[GC Worker End (ms): Min: 150.9, Avg: 150.9, Max: 150.9, Diff: 0.0]
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.2 ms]
[Other: 0.5 ms]
[Choose CSet: 0.0 ms]
[Ref Proc: 0.2 ms]
[Ref Enq: 0.0 ms]
[Redirty Cards: 0.2 ms]
[Humongous Register: 0.0 ms]
[Humongous Reclaim: 0.0 ms]
[Free CSet: 0.0 ms]
[Eden: 12.0M(12.0M)->0.0B(10.0M) Survivors: 0.0B->2048.0K Heap: 12.0M(254.0M)->1120.0K(254.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
0.153: [GC pause (G1 Evacuation Pause) (young), 0.0014333 secs]
[Parallel Time: 1.0 ms, GC Workers: 8]
[GC Worker Start (ms): Min: 152.9, Avg: 152.9, Max: 153.2, Diff: 0.3]
[Ext Root Scanning (ms): Min: 0.0, Avg: 0.1, Max: 0.2, Diff: 0.2, Sum: 1.1]
[Update RS (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Processed Buffers: Min: 0, Avg: 0.0, Max: 0, Diff: 0, Sum: 0]
[Scan RS (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Code Root Scanning (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Object Copy (ms): Min: 0.5, Avg: 0.6, Max: 0.7, Diff: 0.1, Sum: 5.1]
[Termination (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.1]
[Termination Attempts: Min: 1, Avg: 1.9, Max: 4, Diff: 3, Sum: 15]
[GC Worker Other (ms): Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[GC Worker Total (ms): Min: 0.6, Avg: 0.8, Max: 0.9, Diff: 0.3, Sum: 6.4]
[GC Worker End (ms): Min: 153.7, Avg: 153.7, Max: 153.7, Diff: 0.0]
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.1 ms]
[Other: 0.3 ms]
[Choose CSet: 0.0 ms]
[Ref Proc: 0.2 ms]
[Ref Enq: 0.0 ms]
[Redirty Cards: 0.1 ms]
[Humongous Register: 0.0 ms]
[Humongous Reclaim: 0.0 ms]
[Free CSet: 0.0 ms]
[Eden: 10.0M(10.0M)->0.0B(10.0M) Survivors: 2048.0K->2048.0K Heap: 11.1M(254.0M)->3666.5K(254.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
Heap
garbage-first heap total 260096K, used 7762K [0x00000000f0000000, 0x00000000f01007f0, 0x0000000100000000)
region size 1024K, 7 young (7168K), 2 survivors (2048K)
Metaspace used 3451K, capacity 4496K, committed 4864K, reserved 1056768K
class space used 376K, capacity 388K, committed 512K, reserved 1048576K
很明显是发生了两次GC的,并且两次GC之后Eden区的Region数量并没有发生变化。
4.打开实验选项查看YGC的详情日志信息
(1)参数设置
(2)模拟代码
(3)代码运行相关的GC日志
(4)问题
(1)参数设置
具体参数如下:
-Xmx256M -XX:+UseG1GC
-XX:+PrintGCTimeStamps -XX:+PrintGCDetails
-XX:+UnlockExperimentalVMOptions -XX:G1LogLevel=finest
-XX:MaxGCPauseMillis=20 -Xloggc:gc.log
在参数上没有什么特别的变化,主要是添加了两个参数。一个是开启实验选项,用于设置日志的输出级别,日志的输出级别设置为-XX:G1LogLevel=finest
(2)模拟代码
public class YGCDemo1FinestLog {
//5% * 256 = 12M左右
//因为256M / 2048小于1 ,根据G1的处理,会把一个Region定为1M
public static void main(String[] args) throws Exception{
//设定一个对象为256K,如果太大,会直接走大对象分配无法模拟YGC
byte[] data = new byte[1024 * 256];
for (int i = 0; i< 36; i++) {
data = new byte[1024 * 256];
}
//创建n个对象,加入到一个list中,不回收这些数组
ArrayList list = new ArrayList();
for (int i = 0; i<= 36; i++) {
byte[] data2 = new byte[1024 * 256];
//一部分字节数组加入到list中
if (i % 3 == 0) {
list.add(data2);
byte[] bigData1 = new byte[1024 * 512];
byte[] bigData2 = new byte[1024 * 1024];
}
//随机某次,我们把list直接清空掉,最终可能会留下一些存活对象
//这块儿主要是模拟程序正常运行过程中对象可能会被回收
if (i % Math.random() * 10 == 0) {
list.clear();
}
}
}
}
接下来解析一下代码:
一.首先第一段创建了一个data字节数组,大小是1024 * 256 = 256K
因为对象如果超过了Region大小,是一定会被当作大对象单独存储的,所以要把这个对象的大小设置的合理一点。而对于我们设置的这一套参数,堆内存是256M。通过计算不难发现,Region的大小是1M。因为256M/ 2048很显然小于1M,然后会自动调整,调整成1M大小。那么将对象设置成256K大小,就能正常分配到新生代Region中。
二.然后使用了一个循环,在循环中不断去创建新的数组
并且循环的次数设置在了36,为什么设置成36呢?原因是,在G1启动时会初始化新生代的大小。新生代的大小占总空间的5%,也就是256M * 5% 约等于12M。也就是说,当程序用掉了12个Region时,就会触发第一次YGC。
三.在触发了YGC之后
此时data这个变量还在,所以还是会有一个256K的字节数组存活。加上一些堆内存中存在的一些未知对象也可能会存活,那么在第一次回收后,应该会有不到1M的对象存活下来进入Survivor区域。
四.保留一些存活对象到list中
//创建n个对象,加入到一个list中,不回收这些数组
ArrayList list = new ArrayList();
for (int i = 0; i<= 36; i++) {
byte[] data2 = new byte[1024 * 256];
//一部分字节数组加入到list中
if (i % 3 == 0) {
list.add(data2);
byte[] bigData1 = new byte[1024 * 512];
byte[] bigData2 = new byte[1024 * 1024];
}
...
}
代码运行中,这个地方会模拟出每3次执行就放一点存活对象到数组中。
五.紧接着,随机清理一下list,模拟每次运行垃圾和存活对象不定的情况
(3)代码运行相关的GC日志
//这些内容是我们设置的启动参数,其中需要注意的是,最大停顿时间,和堆内存的大小;
CommandLine flags: -XX:G1LogLevel=finest -XX:InitialHeapSize=266064256 -XX:MaxGCPauseMillis=20 -XX:MaxHeapSize=268435456 -XX:+PrintGC -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+UnlockExperimentalVMOptions -XX:+UseCompressedClassPointers -XX:+UseCompressedOops -XX:+UseG1GC -XX:-UseLargePagesIndividualAllocation
//接下来看这里的内容,GC线程还是8个,本次GC造成的停顿时间是0.00178,很明显是在20ms内的;
0.189: [GC pause (G1 Evacuation Pause) (young), 0.0020256 secs]
[Parallel Time: 1.4 ms, GC Workers: 8]
//注意,下面这一行信息是说我们这8个线程开始的时间,min表示最早开始的线程时间,avg表示平均开始时间,max表示最晚开始时间,diff表示最早和最晚的时间差,这个diff越大就说明线程启动时间越不均衡,一般这个地方的时间不会差别很大;
[GC Worker Start (ms): 188.7 188.7 188.7 188.7 188.8 188.8 188.8 188.8
Min: 188.7, Avg: 188.7, Max: 188.8, Diff: 0.1]
//8个线程对根(GC Roots)扫描处理的时间,可以注意到,根的类型非常多。除了经常说的Java静态变量、方法局部变量外还有很多其他的一些根
[Ext Root Scanning (ms): 0.3 1.2 0.2 0.0 0.2 0.2 0.2 0.2
Min: 0.0, Avg: 0.3, Max: 1.2, Diff: 1.2, Sum: 2.5]
[Thread Roots (ms): 0.0 1.2 0.1 0.0 0.0 0.0 0.1 0.0
Min: 0.0, Avg: 0.2, Max: 1.2, Diff: 1.2, Sum: 1.5]
[StringTable Roots (ms): 0.0 0.0 0.1 0.0 0.1 0.1 0.1 0.2
Min: 0.0, Avg: 0.1, Max: 0.2, Diff: 0.2, Sum: 0.6]
[Universe Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[JNI Handles Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[ObjectSynchronizer Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[FlatProfiler Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Management Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[SystemDictionary Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[CLDG Roots (ms): 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.3, Diff: 0.3, Sum: 0.3]
[JVMTI Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[CodeCache Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[CM RefProcessor Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Wait For Strong CLD (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Weak CLD Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[SATB Filtering (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
//更新RSet的操作,前面提过要先更新RSet后,再进行RSet的扫描,其实在这个日志里就能看的出来;
//对象引用关系更新之后,要通过Refine线程 + 系统程序的线程 + DCQ更新RSet
//下面显示的是GC线程时间都为0.0,说明这块更新没问题
//因为引用更新导致RSet更新的这些操作,都已经由Refine线程提前处理完了;
[Update RS (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Processed Buffers: 0 0 0 0 0 0 0 0
Min: 0, Avg: 0.0, Max: 0, Diff: 0, Sum: 0]
//扫描RSet,处理存活对象;
[Scan RS (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Code Root Scanning (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
//这一步很关键,对象拷贝,就是把存活对象拷贝到一个新的分区,可以看出来,拷贝消耗的时间非常长;
//注意:对象拷贝消耗的时间占了GC绝大部分的时间
[Object Copy (ms): 0.5 0.0 0.6 0.0 0.7 0.6 0.6 0.6
Min: 0.0, Avg: 0.5, Max: 0.7, Diff: 0.7, Sum: 3.8]
//GC线程关闭的一些信息
[Termination (ms): 0.4 0.0 0.4 0.4 0.4 0.4 0.4 0.4
Min: 0.0, Avg: 0.4, Max: 0.4, Diff: 0.4, Sum: 2.8]
[Termination Attempts: 3 1 12 1 16 9 11 9
Min: 1, Avg: 7.8, Max: 16, Diff: 15, Sum: 62]
//GC线程的一些其他工作的处理时间,通常是JVM析构释放资源
[GC Worker Other (ms): 0.0 0.1 0.1 0.9 0.1 0.1 0.0 0.0
Min: 0.0, Avg: 0.2, Max: 0.9, Diff: 0.9, Sum: 1.3]
//并行GC过程花费的总体时间;
[GC Worker Total (ms): 1.3 1.3 1.3 1.3 1.3 1.3 1.3 1.3
Min: 1.3, Avg: 1.3, Max: 1.3, Diff: 0.1, Sum: 10.4]
[GC Worker End (ms): 190.0 190.0 190.0 190.1 190.0 190.1 190.0 190.1
Min: 190.0, Avg: 190.0, Max: 190.1, Diff: 0.0]
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
//清理卡表
[Clear CT: 0.2 ms]
[Other: 0.5 ms]
//选择Cset时间,对YGC来说,一般都是0,因为不需要选
[Choose CSet: 0.0 ms]
[Ref Proc: 0.2 ms]
[Ref Enq: 0.0 ms]
//重构RSet花费的时间
[Redirty Cards: 0.2 ms]
[Parallel Redirty: 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Redirtied Cards: 0 0 0 0 0 0 0 0
Min: 0, Avg: 0.0, Max: 0, Diff: 0, Sum: 0]
//大对象的处理时间
[Humongous Register: 0.0 ms]
[Humongous Total: 0]
[Humongous Candidate: 0]
//如果大对象要回收,回收花费的时间以及回收了多少个大对象
[Humongous Reclaim: 0.0 ms]
[Humongous Reclaimed: 0]
//释放CSet中的分区花费的时间
[Free CSet: 0.0 ms]
[Young Free CSet: 0.0 ms]
[Non-Young Free CSet: 0.0 ms]
//本次回收之后,Eden区和Survivor区的状态
[Eden: 12.0M(12.0M)->0.0B(10.0M) Survivors: 0.0B->2048.0K Heap: 12.0M(254.0M)->1120.0K(254.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
0.199: [GC pause (G1 Evacuation Pause) (young), 0.0015365 secs]
[Parallel Time: 1.0 ms, GC Workers: 8]
[GC Worker Start (ms): 199.4 199.4 199.4 199.4 199.9 199.9 199.9 200.3
Min: 199.4, Avg: 199.7, Max: 200.3, Diff: 0.9]
[Ext Root Scanning (ms): 0.2 0.2 0.2 0.2 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.1, Max: 0.2, Diff: 0.2, Sum: 0.7]
[Thread Roots (ms): 0.0 0.1 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.1, Diff: 0.1, Sum: 0.1]
[StringTable Roots (ms): 0.0 0.1 0.2 0.1 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.1, Max: 0.2, Diff: 0.2, Sum: 0.4]
[Universe Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[JNI Handles Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[ObjectSynchronizer Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[FlatProfiler Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Management Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[SystemDictionary Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[CLDG Roots (ms): 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.2, Diff: 0.2, Sum: 0.2]
[JVMTI Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[CodeCache Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[CM RefProcessor Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Wait For Strong CLD (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Weak CLD Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[SATB Filtering (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Update RS (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Processed Buffers: 0 0 0 0 0 0 0 0
Min: 0, Avg: 0.0, Max: 0, Diff: 0, Sum: 0]
[Scan RS (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Code Root Scanning (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Object Copy (ms): 0.6 0.7 0.7 0.7 0.4 0.4 0.3 0.0
Min: 0.0, Avg: 0.5, Max: 0.7, Diff: 0.7, Sum: 3.7]
[Termination (ms): 0.1 0.0 0.1 0.1 0.1 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.1, Diff: 0.1, Sum: 0.4]
[Termination Attempts: 1 1 5 1 4 1 1 1
Min: 1, Avg: 1.9, Max: 5, Diff: 4, Sum: 15]
[GC Worker Other (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[GC Worker Total (ms): 0.9 0.9 0.9 0.9 0.4 0.4 0.4 0.0
Min: 0.0, Avg: 0.6, Max: 0.9, Diff: 0.9, Sum: 4.9]
[GC Worker End (ms): 200.3 200.3 200.3 200.3 200.3 200.3 200.3 200.3
Min: 200.3, Avg: 200.3, Max: 200.3, Diff: 0.0]
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.2 ms]
[Other: 0.4 ms]
[Choose CSet: 0.0 ms]
[Ref Proc: 0.1 ms]
[Ref Enq: 0.0 ms]
[Redirty Cards: 0.2 ms]
[Parallel Redirty: 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Redirtied Cards: 67 0 0 0 0 0 0 0
Min: 0, Avg: 8.4, Max: 67, Diff: 67, Sum: 67]
[Humongous Register: 0.0 ms]
[Humongous Total: 16]
[Humongous Candidate: 16]
[Humongous Reclaim: 0.0 ms]
[Humongous Reclaimed: 16]
[Free CSet: 0.0 ms]
[Young Free CSet: 0.0 ms]
[Non-Young Free CSet: 0.0 ms]
[Eden: 10.0M(10.0M)->0.0B(10.0M) Survivors: 2048.0K->2048.0K Heap: 23.1M(254.0M)->3357.5K(254.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//以下这个就是在我们程序运行结束之后,堆内存的使用状态
Heap
garbage-first heap total 260096K, used 15133K [0x00000000f0000000, 0x00000000f01007f0, 0x0000000100000000)
region size 1024K, 7 young (7168K), 2 survivors (2048K)
Metaspace used 3465K, capacity 4496K, committed 4864K, reserved 1056768K
class space used 377K, capacity 388K, committed 512K, reserved 1048576K
(4)问题
模拟YGC一直运行的日志输出情况,把代码改造一下,加一个while循环,参数不变:
-Xmx256M -XX:+UseG1GC
-XX:+PrintGCTimeStamps -XX:+PrintGCDetails
-XX:+UnlockExperimentalVMOptions -XX:G1LogLevel=finest
-XX:MaxGCPauseMillis=20 -Xloggc:gc.log
如下这样的代码,如果一直运行下去,新生代会发生怎样的变化?结合TLAB的变化、预测模型、YGC后的新生代动态调整、以及新生代动态调整的比例,来思考下YGC在程序持续运行一段时间后,会发生怎样的变化?
//5% * 256 = 12M左右
//因为256M / 2048小于1,根据G1的处理,会把一个Region定为1M
public static void main(String[] args) throws Exception {
while(true) {
//设定一个对象为256K,如果太大,会直接走大对象分配无法模拟YGC
byte[] data = new byte[1024 * 256];
for (int i = 0; i< 36; i++) {
data = new byte[1024 * 256];
}
//创建n个对象,加入到一个list中,不回收这些数组
ArrayList list = new ArrayList();
for (int i = 0; i<= 36; i++) {
byte[] data2 = new byte[1024 * 256];
//一部分字节数组加入到list中
if (i % 3 == 0) {
list.add(data2);
byte[] bigData1 = new byte[1024 * 512];
byte[] bigData2 = new byte[1024 * 1024];
}
//随机某次把list直接清空掉,最终可能会留下一些存活对象
//这块主要是模拟程序正常运行过程中对象可能会被回收
if (i % Math.random()*10 == 0) {
list.clear();
}
}
}
}
5.Mixed GC日志信息之初始标记过程
(1)参数设置
(2)模拟代码
(3)代码运行相关的GC日志
(1)参数设置
具体参数如下:
-Xmx256M -XX:+UseG1GC
-XX:+PrintGCTimeStamps -XX:+PrintGCDetails
-XX:+UnlockExperimentalVMOptions
-XX:G1LogLevel=finest -XX:MaxGCPauseMillis=20 -Xloggc:gc.log
在参数上没有什么特别的变化,主要是添加了两个参数。一个是开启实验选项,用于设置日志的输出级别,日志的输出级别设置为-XX:G1LogLevel=finest
(2)模拟代码
private static final ArrayList list = new ArrayList();
//5% * 256 = 12M左右
//因为256M / 2048小于1,根据G1的处理,会把一个Region定为1M
public static void main(String[] args) throws Exception {
while(true) {
//设定一个对象为256K,如果太大,会直接走大对象分配无法模拟YGC
byte[] data = new byte[1024 * 256];
for (int i = 0; i< 36; i++) {
data = new byte[1024 * 256];
}
//创建n个对象,加入到一个list中,不回收这些数组
for (int i = 0; i<= 36; i++) {
byte[] data2 = new byte[1024 * 256];
//一部分字节数组加入到list中
if (i % 2 == 0) {
list.add(data2);
byte[] bigData1 = new byte[1024 * 512];
}
}
Thread.sleep(1000);
}
}
下面来解析一下代码:
一.首先第一段创建了一个data字节数组,大小是1024 * 256 = 256KB
因为对象如果超过了Region大小一半,是一定会被当作大对象单独存储的,所以要把这个对象的大小设置的合理一点。而对于设置的参数,堆内存是256M。通过计算不难发现,Region的大小是1M。因为256M / 2048很显然小于1M,然后会自动调整,调整成1M大小。那么将对象设置成256K大小,就能正常分配到新生代Region中。
二.然后使用了一个循环,在循环中会不断创建新的数组
并且循环的次数设置在了36,为什么设置成36呢?原因是,在G1启动时会初始化新生代的大小。新生代的大小占总空间的5%,也就是256M * 5%约等于12M。也就是用掉12个Region时,就会触发第一次YGC。
三.在触发了YGC之后
由于data这个变量还在,所以还是会有一个256K的字节数组存活。加上一些堆内存中存在的一些未知对象也可能会存活,那么在第一次回收后,应该会有不到1M的对象存活下来进入Survivor区域。
四.比较大的改变在这里:
//一部分字节数组加入到list中
if (i % 2 == 0) {
list.add(data2);
byte[] bigData1 = new byte[1024 * 512];
}
代码运行中,这个地方模拟出,每2次执行就把存活对象放入到数组中。这样做是为了更快触发Mixed GC,因为Mixed GC的触发条件是老年代使用了45%,并且垃圾清理的过程并不一定是跟在Initial Mark之后的。
也就是说,即使触发了初始标记过程,也不一定会启动Mixed GC。所以要让堆内存里的存活对象始终在增长,保证它一定会触发Mixed GC。
(3)代码运行相关的GC日志
从如下这个日志可以看出:首先会触发了一次YGC,并且时间并不长。然后经过很多次YGC后才会触发Mixed GC的初始标记,并且触发Mixed GC的初始标记后,紧接着的也不一定就是Mixed GC。因为这期间应用程序还会运行,就有可能继续触发YGC。
0.177: [GC pause (G1 Evacuation Pause) (young), 0.0017408 secs]
// 启动了8个线程去执行,并且并行的执行时间1.2ms
[Parallel Time: 1.2 ms, GC Workers: 8]
[GC Worker Start (ms): 177.2 177.2 177.2 177.2 177.2 177.6 178.3 178.3
Min: 177.2, Avg: 177.5, Max: 178.3, Diff: 1.1]
...
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.1 ms]
[Other: 0.4 ms]
...
//从这里可以看出,这次GC是一次非常常规的YGC,GC之前,S区没有,GC之后,下一次S区是1个,也就是1M的空间;
//堆内存在回收前是6144K垃圾,回收后是1048K垃圾
[Eden: 6144.0K(6144.0K)->0.0B(5120.0K) Survivors: 0.0B->1024.0K Heap: 6144.0K(128.0M)->1048.0K(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//第二次回收,我们着重看一下对象的增涨速度
0.180: [GC pause (G1 Evacuation Pause) (young), 0.0016041 secs]
[Parallel Time: 1.1 ms, GC Workers: 8]
[GC Worker Start (ms): 180.3 180.3 180.3 180.3 180.3 180.3 180.6 181.2
Min: 180.3, Avg: 180.4, Max: 181.2, Diff: 1.0]
...
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.1 ms]
[Other: 0.4 ms]
...
//从下面这里可以看到:
//对象增涨的速度其实并不快,并且Eden区的个数,调大了一个
[Eden: 5120.0K(5120.0K)->0.0B(6144.0K) Survivors: 1024.0K->1024.0K Heap: 6168.0K(128.0M)->1427.9K(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//接下来省略很多次普通的YGC,YGC的过程没必要放出来了,因为太多了;
...
...
...
...
//到了这次YGC,依然是很正常的YGC,还是常规的YGC结果;
7.289: [GC pause (G1 Evacuation Pause) (young), 0.0029624 secs]
[Parallel Time: 1.8 ms, GC Workers: 8]
[GC Worker Start (ms): 7288.9 7288.9 7288.9 7289.0 7289.0 7289.0 7289.0 7289.0
Min: 7288.9, Avg: 7288.9, Max: 7289.0, Diff: 0.1]
...
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.5 ms]
[Other: 0.6 ms]
...
//注意这里,此时堆内存的使用已经非常多了;
//Eden区在回收以后是0B,S区扩展到了6M左右,堆内存总使用量47.5M,此时这个使用量其实已经比较高了;
[Eden: 38.0M(38.0M)->0.0B(57.0M) Survivors: 5120.0K->6144.0K Heap: 89.3M(128.0M)->47.5M(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//重点看这里,发现已经开启了初始标记(可见此时也还是一次YGC);
8.309: [GC pause (G1 Humongous Allocation) (young) (initial-mark), 0.0014780 secs]
[Parallel Time: 1.0 ms, GC Workers: 8]
[GC Worker Start (ms): 8308.9 8308.9 8308.9 8308.9 8308.9 8308.9 8309.0 8309.1
Min: 8308.9, Avg: 8309.0, Max: 8309.1, Diff: 0.2]
...
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.1 ms]
[Other: 0.3 ms]
...
//从这次回收之后的数据可以看到,确实是遵循45%这个规则;
//YGC垃圾回收后,即将分配的对象 + 老年代的所有对象占用的空间达到了45%时,就会触发Mixed GC,也就是开启并发标记;
[Eden: 24.0M(57.0M)->0.0B(49.0M) Survivors: 6144.0K->6144.0K Heap: 78.8M(128.0M)->53.0M(128.0M)]
[Times: user=0.09 sys=0.00, real=0.00 secs]
//并发根扫描的开始和结束
8.311: [GC concurrent-root-region-scan-start]
8.311: [GC concurrent-root-region-scan-end, 0.0000467 secs]
//并发标记的开始和结束
8.311: [GC concurrent-mark-start]
8.313: [GC concurrent-mark-end, 0.0018556 secs]
//重新标记阶段
8.313: [GC remark 8.313: [Finalize Marking, 0.0001905 secs] 8.313: [GC ref-proc, 0.0003040 secs] 8.313: [Unloading 8.313: [System Dictionary Unloading, 0.0000130 secs] 8.313: [Parallel Unloading, 0.0005763 secs] 8.314: [Deallocate Metadata, 0.0000144 secs], 0.0006416 secs], 0.0014267 secs]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//预清理阶段
8.314: [GC cleanup 75M->75M(128M), 0.0016325 secs]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//又进行了一次YGC,因为Mixed GC初始标记后的并发标记阶段会和系统程序一起并发运行,所以这里是系统程序触发的YGC
9.331: [GC pause (G1 Evacuation Pause) (young) (to-space exhausted), 0.0008772 secs]
[Parallel Time: 0.4 ms, GC Workers: 8]
[GC Worker Start (ms): 9331.0 9331.0 9331.0 9331.0 9331.0 9331.0 9331.0 9331.0
Min: 9331.0, Avg: 9331.0, Max: 9331.0, Diff: 0.1]
...
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.1 ms]
[Other: 0.4 ms]
...
[Eden: 37.0M(49.0M)->0.0B(6144.0K) Survivors: 6144.0K->0.0B Heap: 108.5M(128.0M)->78.8M(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//接下来才是打印触发的Mixed GC
10.347: [GC pause (G1 Evacuation Pause) (mixed), 0.0017163 secs]
[Parallel Time: 1.2 ms, GC Workers: 8]
[GC Worker Start (ms): 10347.3 10347.3 10347.3 10347.3 10347.3 10347.3 10347.5 10347.6
Min: 10347.3, Avg: 10347.4, Max: 10347.6, Diff: 0.3]
...
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.1 ms]
[Other: 0.4 ms]
...
[Eden: 6144.0K(6144.0K)->0.0B(5120.0K) Survivors: 0.0B->1024.0K Heap: 85.3M(128.0M)->78.8M(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
10.350: [GC pause (G1 Evacuation Pause) (mixed), 0.0015104 secs]
[Parallel Time: 0.8 ms, GC Workers: 8]
[GC Worker Start (ms): 10350.1 10350.2 10350.2 10350.2 10350.2 10350.2 10350.2 10350.3
Min: 10350.1, Avg: 10350.2, Max: 10350.3, Diff: 0.2]
...
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.1 ms]
[Other: 0.5 ms]
...
[Eden: 5120.0K(5120.0K)->0.0B(27.0M) Survivors: 1024.0K->1024.0K Heap: 83.8M(128.0M)->79.5M(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//又出现了一次Mixed GC的初始标记
10.353: [GC pause (G1 Humongous Allocation) (young) (initial-mark), 0.0009719 secs]
[Parallel Time: 0.4 ms, GC Workers: 8]
[GC Worker Start (ms): 10352.6 10352.6 10352.7 10352.7 10352.7 10352.7 10352.7 10352.7
Min: 10352.6, Avg: 10352.7, Max: 10352.7, Diff: 0.1]
...
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.2 ms]
[Other: 0.4 ms]
...
[Eden: 3072.0K(27.0M)->0.0B(28.0M) Survivors: 1024.0K->1024.0K Heap: 82.0M(128.0M)->79.8M(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
10.354: [GC concurrent-root-region-scan-start]
10.354: [GC concurrent-root-region-scan-end, 0.0000481 secs]
10.354: [GC concurrent-mark-start]
10.356: [GC concurrent-mark-end, 0.0019111 secs]
10.356: [GC remark 10.356: [Finalize Marking, 0.0002410 secs] 10.357: [GC ref-proc, 0.0002787 secs] 10.357: [Unloading 10.357: [System Dictionary Unloading, 0.0000089 secs] 10.357: [Parallel Unloading, 0.0005675 secs] 10.357: [Deallocate Metadata, 0.0000145 secs], 0.0006318 secs], 0.0014424 secs]
[Times: user=0.01 sys=0.00, real=0.00 secs]
10.358: [GC cleanup 101M->101M(128M), 0.0006501 secs]
[Times: user=0.00 sys=0.00, real=0.00 secs]
6.Mixed GC日志信息之混合回收过程
继续前面的Mixed GC日志分析:
...
//到了这次YGC,依然是很正常的YGC;
7.289: [GC pause (G1 Evacuation Pause) (young), 0.0029624 secs]
[Parallel Time: 1.8 ms, GC Workers: 8]
...
//注意,此时堆内存的使用已经非常多了;
//Eden区在回收以后,是0B,S区扩展到了6M左右,堆内存总使用量47.5M,此时这个使用量其实已经比较高了;
[Eden: 38.0M(38.0M)->0.0B(57.0M) Survivors: 5120.0K->6144.0K Heap: 89.3M(128.0M)->47.5M(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//接下来这里,发现已经开启了初始标记;
8.309: [GC pause (G1 Humongous Allocation) (young) (initial-mark), 0.0014780 secs]
[Parallel Time: 1.0 ms, GC Workers: 8]
...
//这次回收之后的数据如下一行所示;
//从数据中可以看到,确实是遵循45%这个规则来的;
//垃圾回收后,即将分配的对象 + 老年代的所有对象占用的空间达到了45%的时候,就会触发Mixed GC,也就是开启并发标记;
[Eden: 24.0M(57.0M)->0.0B(49.0M) Survivors: 6144.0K->6144.0K Heap: 78.8M(128.0M)->53.0M(128.0M)]
[Times: user=0.09 sys=0.00, real=0.00 secs]
//从这里也可以看出来,初始标记状态,其实就是一次YGC,但是需要注意的是,初始标记之后,并不一定会执行Mixed GC
//初始标记之后会进入并发标记;
//下面这几行实际上做的是从YHR中的Survivor分区开始进行并发标记的GC Roots扫描;
//可以理解成,找到Survivor区的所有存活对象,并作为根对象,这一部分还没有对GC Roots直接引用的老年代对象做扫描
8.311: [GC concurrent-root-region-scan-start]
//扫描结束,用时0.0000467秒,这个扫描过程需要注意的是:它是和应用程序同时进行的,并且有多个线程参与
8.311: [GC concurrent-root-region-scan-end, 0.0000467 secs]
//并发标记的全量扫描过程,就是以Survivor + GC Roots(方法栈里面的局部变量,静态变量等),标记整个堆里面的所有存活对象
8.311: [GC concurrent-mark-start]
//并发标记结束,花费了0.0018565秒
8.313: [GC concurrent-mark-end, 0.0018556 secs]
//最终标记阶段,标记程序创建出来的新对象,软引用、弱引用等的处理;
8.313: [GC remark 8.313: [Finalize Marking, 0.0001905 secs] 8.313: [GC ref-proc, 0.0003040 secs] 8.313: [Unloading 8.313: [System Dictionary Unloading, 0.0000130 secs] 8.313: [Parallel Unloading, 0.0005763 secs] 8.314: [Deallocate Metadata, 0.0000144 secs], 0.0006416 secs], 0.0014267 secs]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//预清理处理,这个地方清除的是全部都是垃圾的那些Region分区,从这里的日志可以看出来,没有清理掉任何一个Region
8.314: [GC cleanup 75M->75M(128M), 0.0016325 secs]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//紧接着,又进入了YGC(此时S区不足需要清空S区);
//而且是标明了to-space exhausted,这说明to区在本次GC之后不够用了,要做一些对象升入老年代的操作,调整S区大小
9.331: [GC pause (G1 Evacuation Pause) (young) (to-space exhausted), 0.0008772 secs]
[Parallel Time: 0.4 ms, GC Workers: 8]
[GC Worker Start (ms): 9331.0 9331.0 9331.0 9331.0 9331.0 9331.0 9331.0 9331.0
Min: 9331.0, Avg: 9331.0, Max: 9331.0, Diff: 0.1]
...
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.1 ms]
[Other: 0.4 ms]
...
//这里可以看出来,在本次回收之后,S区已经变成了0了
[Eden: 37.0M(49.0M)->0.0B(6144.0K) Survivors: 6144.0K->0.0B Heap: 108.5M(128.0M)->78.8M(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//然后开始进入混合回收阶段,混合回收的内容和YGC的内容其实差不多
10.347: [GC pause (G1 Evacuation Pause) (mixed), 0.0017163 secs]
//并行时间1.2ms,8个线程参与
[Parallel Time: 1.2 ms, GC Workers: 8]
[GC Worker Start (ms): 10347.3 10347.3 10347.3 10347.3 10347.3 10347.3 10347.5 10347.6
Min: 10347.3, Avg: 10347.4, Max: 10347.6, Diff: 0.3]
[Ext Root Scanning (ms): 0.2 0.2 0.2 0.2 0.2 0.1 0.0 0.0
Min: 0.0, Avg: 0.1, Max: 0.2, Diff: 0.2, Sum: 1.0]
[Thread Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[StringTable Roots (ms): 0.1 0.2 0.1 0.1 0.2 0.1 0.0 0.0
Min: 0.0, Avg: 0.1, Max: 0.2, Diff: 0.2, Sum: 0.8]
[Universe Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[JNI Handles Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[ObjectSynchronizer Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[FlatProfiler Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Management Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[SystemDictionary Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[CLDG Roots (ms): 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.1, Diff: 0.1, Sum: 0.1]
[JVMTI Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[CodeCache Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.1, Diff: 0.1, Sum: 0.2]
[CM RefProcessor Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Wait For Strong CLD (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Weak CLD Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[SATB Filtering (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Update RS (ms): 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.1, Diff: 0.1, Sum: 0.2]
[Processed Buffers: 2 1 1 0 0 1 0 0
Min: 0, Avg: 0.6, Max: 2, Diff: 2, Sum: 5]
[Scan RS (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Code Root Scanning (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Object Copy (ms): 0.8 0.9 0.9 0.8 0.8 0.8 0.8 0.7
Min: 0.7, Avg: 0.8, Max: 0.9, Diff: 0.2, Sum: 6.4]
[Termination (ms): 0.1 0.1 0.0 0.1 0.1 0.1 0.1 0.1
Min: 0.0, Avg: 0.1, Max: 0.1, Diff: 0.1, Sum: 0.7]
[Termination Attempts: 1 1 1 1 1 1 1 1
Min: 1, Avg: 1.0, Max: 1, Diff: 0, Sum: 8]
[GC Worker Other (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[GC Worker Total (ms): 1.1 1.1 1.1 1.1 1.1 1.1 0.9 0.8
Min: 0.8, Avg: 1.0, Max: 1.1, Diff: 0.3, Sum: 8.4]
[GC Worker End (ms): 10348.4 10348.4 10348.4 10348.4 10348.4 10348.4 10348.4 10348.4
Min: 10348.4, Avg: 10348.4, Max: 10348.4, Diff: 0.0]
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.1 ms]
[Other: 0.4 ms]
[Choose CSet: 0.0 ms]
[Ref Proc: 0.2 ms]
[Ref Enq: 0.0 ms]
[Redirty Cards: 0.1 ms]
[Parallel Redirty: 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Redirtied Cards: 4 0 0 0 0 0 0 0
Min: 0, Avg: 0.5, Max: 4, Diff: 4, Sum: 4]
[Humongous Register: 0.0 ms]
[Humongous Total: 1]
[Humongous Candidate: 1]
[Humongous Reclaim: 0.0 ms]
[Humongous Reclaimed: 1]
[Free CSet: 0.0 ms]
[Young Free CSet: 0.0 ms]
[Non-Young Free CSet: 0.0 ms]
[Eden: 6144.0K(6144.0K)->0.0B(5120.0K) Survivors: 0.0B->1024.0K Heap: 85.3M(128.0M)->78.8M(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//由上述可见,Mixed GC的回收阶段的过程和YGC几乎一模一样,什么原因呢?
//其实就是复用了代码,然后主要的区别在于选择的Cset不同
//下面又是一次Mixed GC,因为会分批次,默认最多分8次
10.350: [GC pause (G1 Evacuation Pause) (mixed), 0.0015104 secs]
[Parallel Time: 0.8 ms, GC Workers: 8]
[GC Worker Start (ms): 10350.1 10350.2 10350.2 10350.2 10350.2 10350.2 10350.2 10350.3
Min: 10350.1, Avg: 10350.2, Max: 10350.3, Diff: 0.2]
...
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.1 ms]
[Other: 0.5 ms]
...
[Eden: 5120.0K(5120.0K)->0.0B(27.0M) Survivors: 1024.0K->1024.0K Heap: 83.8M(128.0M)->79.5M(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//又一次触发了初始标记
//其实看到这里触发初始标记,也可以大概猜到,后续的Mixed GC也会非常频繁,并且初始标记也会经常发生,因为老年代的存活对象太多了;
10.353: [GC pause (G1 Humongous Allocation) (young) (initial-mark), 0.0009719 secs]
[Parallel Time: 0.4 ms, GC Workers: 8]
[GC Worker Start (ms): 10352.6 10352.6 10352.7 10352.7 10352.7 10352.7 10352.7 10352.7
Min: 10352.6, Avg: 10352.7, Max: 10352.7, Diff: 0.1]
...
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.2 ms]
[Other: 0.4 ms]
...
[Eden: 3072.0K(27.0M)->0.0B(28.0M) Survivors: 1024.0K->1024.0K Heap: 82.0M(128.0M)->79.8M(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
10.354: [GC concurrent-root-region-scan-start]
10.354: [GC concurrent-root-region-scan-end, 0.0000481 secs]
10.354: [GC concurrent-mark-start]
10.356: [GC concurrent-mark-end, 0.0019111 secs]
10.356: [GC remark 10.356: [Finalize Marking, 0.0002410 secs] 10.357: [GC ref-proc, 0.0002787 secs] 10.357: [Unloading 10.357: [System Dictionary Unloading, 0.0000089 secs] 10.357: [Parallel Unloading, 0.0005675 secs] 10.357: [Deallocate Metadata, 0.0000145 secs], 0.0006318 secs], 0.0014424 secs]
[Times: user=0.01 sys=0.00, real=0.00 secs]
10.358: [GC cleanup 101M->101M(128M), 0.0006501 secs]
[Times: user=0.00 sys=0.00, real=0.00 secs]
Mixed GC的过程相对复杂。在触发了初始标记以后,并不一定紧跟的就是Mixed GC的回收阶段。有可能会继续进行YGC的过程,然后才进入Mixd GC的回收阶段。
由此可见,对于Mixed GC来说:从初始标记到并发标记到回收阶段整个流程,不一定是连续的整个流程。因为在并发标记阶段程序会运行,接下来有可能再触发一次YGC。并且在可回收的垃圾不足5%时,也不会触发Mixed GC的回收阶段。同时如果进行了多次GC后,小于了这个5%,也可能不继续进行混合回收。当然这个5%不太好模拟,如果是线上的程序其实更好模拟一些。
7.Mixed GC日志信息之Region的详细信息和标记过程的详细信息
(1)参数设置
(2)模拟代码
(3)代码运行相关的GC日志
(4)这些数据有什么作用
(1)参数设置
具体参数如下:
-Xmx128M -XX:+UseG1GC
-XX:+PrintGCTimeStamps -XX:+PrintGCDetails
-XX:+UnlockExperimentalVMOptions -XX:G1LogLevel=finest
-XX:+UnlockDiagnosticVMOptions -XX:+G1PrintRegionLivenessInfo
-XX:+G1SummarizeConcMark -XX:MaxGCPauseMillis=20
-Xloggc:gc.log
和前面设置的JVM参数有区别的是如下这几个:
打开数据分析选项:-XX:+UnlockDiagnosticVMOptions
打印所有的region的使用详情信息:-XX:+G1PrintRegionLivenessInfo
打开并发标记的详细信息:-XX:+G1SummarizeConcMark
(2)模拟代码
public static void main(String[] args) throws Exception {
while(true) {
//设定一个对象为256K,如果太大,会直接走大对象分配无法模拟YGC
byte[] data = new byte[1024 * 256];
for (int i = 0; i< 36; i++) {
data = new byte[1024 * 256];
}
//创建n个对象,加入到一个list中,不回收这些数组
for (int i = 0; i<= 36; i++) {
byte[] data2 = new byte[1024 * 256];
//一部分字节数组加入到list中
if (i % 2 == 0) {
list.add(data2);
byte[] bigData1 = new byte[1024 * 512];
}
}
Thread.sleep(100);
}
}
(3)代码运行相关的GC日志
CommandLine flags: -XX:G1LogLevel=finest -XX:+G1PrintRegionLivenessInfo -XX:+G1SummarizeConcMark -XX:InitialHeapSize=134217728 -XX:MaxGCPauseMillis=20 -XX:MaxHeapSize=134217728 -XX:+PrintGC -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+UnlockDiagnosticVMOptions -XX:+UnlockExperimentalVMOptions -XX:+UseCompressedClassPointers -XX:+UseCompressedOops -XX:+UseG1GC -XX:-UseLargePagesIndividualAllocation
0.130: [GC pause (G1 Humongous Allocation) (young) (initial-mark), 0.0020150 secs]
[Parallel Time: 1.0 ms, GC Workers: 8]
[GC Worker Start (ms): 129.7 129.8 129.8 129.8 129.8 129.8 129.8 129.8
Min: 129.7, Avg: 129.8, Max: 129.8, Diff: 0.1]
...
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
[Clear CT: 0.4 ms]
[Other: 0.7 ms]
...
[Eden: 3072.0K(6144.0K)->0.0B(5120.0K) Survivors: 0.0B->1024.0K Heap: 31.1M(128.0M)->1848.1K(128.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
0.132: [GC concurrent-root-region-scan-start]
0.133: [GC concurrent-root-region-scan-end, 0.0008120 secs]
0.133: [GC concurrent-mark-start]
0.133: [GC concurrent-mark-end, 0.0001406 secs]
0.133: [GC remark 0.133: [Finalize Marking, 0.0004982 secs] 0.134: [GC ref-proc, 0.0002247 secs] 0.134: [Unloading 0.134: [System Dictionary Unloading, 0.0000091 secs] 0.134: [Parallel Unloading, 0.0005040 secs] 0.134: [Deallocate Metadata, 0.0000141 secs], 0.0005701 secs], 0.0016329 secs]
[Times: user=0.00 sys=0.00, real=0.00 secs]
0.135: [GC cleanup
### PHASE Post-Marking @ 0.135
### HEAP reserved: 0x00000000f8000000-0x0000000100000000 region-size: 1048576
### type(分区类型) address-range(地址范围) used(使用量) prev-live next-live gc-eff(GC效率) remset(Rset大小) code-roots
### (bytes) (bytes) (bytes) (bytes/ms) (bytes) (bytes)
### HUMS 0x00000000f8000000-0x00000000f8100000 1048576 1048576 1048576 0.0 2904 16
### HUMC 0x00000000f8100000-0x00000000f8200000 16 16 16 0.0 2904 16
...
### HUMS 0x00000000fae00000-0x00000000faf00000 1048576 1048576 1048576 0.0 2904 16
### HUMC 0x00000000faf00000-0x00000000fb000000 16 16 16 0.0 2904 16
### FREE 0x00000000fb000000-0x00000000fb100000 0 0 0 0.0 2904 16
### FREE 0x00000000fb100000-0x00000000fb200000 0 0 0 0.0 2904 16
### FREE 0x00000000fb200000-0x00000000fb300000 0 0 0 0.0 2904 16
### FREE 0x00000000fb300000-0x00000000fb400000 0 0 0 0.0 2904 16
### FREE 0x00000000fb400000-0x00000000fb500000 0 0 0 0.0 2904 16
### FREE 0x00000000fb500000-0x00000000fb600000 0 0 0 0.0 2904 16
### FREE 0x00000000fb600000-0x00000000fb700000 0 0 0 0.0 2904 16
### FREE 0x00000000fb700000-0x00000000fb800000 0 0 0 0.0 2904 16
### HUMS 0x00000000fb800000-0x00000000fb900000 1048576 1048576 1048576 0.0 2904 16
### HUMC 0x00000000fb900000-0x00000000fba00000 16 16 16 0.0 2904 16
### FREE 0x00000000fba00000-0x00000000fbb00000 0 0 0 0.0 2904 16
...
### FREE 0x00000000ffa00000-0x00000000ffb00000 0 0 0 0.0 2904 16
### FREE 0x00000000ffb00000-0x00000000ffc00000 0 0 0 0.0 2904 16
### SURV 0x00000000ffc00000-0x00000000ffd00000 843840 843840 843840 0.0 3056 168
### FREE 0x00000000ffd00000-0x00000000ffe00000 0 0 0 0.0 2904 16
### FREE 0x00000000ffe00000-0x00000000fff00000 0 0 0 0.0 2904 16
### EDEN 0x00000000fff00000-0x0000000100000000 104904 104904 104904 0.0 2904 16
###
### SUMMARY capacity: 128.00 MB used: 25.91 MB / 20.24 % prev-live: 25.91 MB / 20.24 % next-live: 25.91 MB / 20.24 % remset: 0.36 MB code-roots: 0.00 MB
### PHASE Post-Sorting @ 0.136
### HEAP reserved: 0x00000000f8000000-0x0000000100000000 region-size: 1048576
###
### type address-range used prev-live next-live gc-eff remset code-roots
### (bytes) (bytes) (bytes) (bytes/ms) (bytes) (bytes)
###
### SUMMARY capacity: 0.00 MB used: 0.00 MB / 0.00 % prev-live: 0.00 MB / 0.00 % next-live: 0.00 MB / 0.00 % remset: 0.01 MB code-roots: 0.00 MB
25M->25M(128M), 0.0010444 secs]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//堆内存退出前的使用状态
Heap
garbage-first heap total 131072K, used 65513K [0x00000000f8000000, 0x00000000f8100400, 0x0000000100000000)
region size 1024K, 1 young (1024K), 0 survivors (0K)
Metaspace used 3969K, capacity 4568K, committed 4864K, reserved 1056768K
class space used 437K, capacity 460K, committed 512K, reserved 1048576K
//并发标记的状态,其中包括标记类型,比如初始标记状态、init marks、并发标记remarks、最终标记final marks、weak refs,弱引用处理
Concurrent marking:
0 init marks: total time = 0.00 s (avg = 0.00 ms).
2 remarks: total time = 0.00 s (avg = 1.36 ms).
[std. dev = 0.00 ms, max = 1.36 ms]
2 final marks: total time = 0.00 s (avg = 0.21 ms).
[std. dev = 0.02 ms, max = 0.23 ms]
2 weak refs: total time = 0.00 s (avg = 1.14 ms).
[std. dev = 0.02 ms, max = 1.16 ms]
2 cleanups: total time = 0.00 s (avg = 0.78 ms).
[std. dev = 0.33 ms, max = 1.11 ms]
//对对象进行计数统计花费的时间
Final counting total time = 0.00 s (avg = 0.24 ms).
//RSet的处理时长
RS scrub total time = 0.00 s (avg = 0.22 ms).
//所有STW的阶段加起来一共停顿了多久
Total stop_world time = 0.00 s.
//所有GC线程在整个并发标记过程中的总花费时间(多个线程并发,每个线程的耗时求和)
Total concurrent time = 0.03 s ( 0.03 s marking).
(4)这些数据有什么作用
平常的话, 如下这几个参数是不需要打开的:
-XX:+UnlockExperimentalVMOptions -XX:G1LogLevel=finest
-XX:+UnlockDiagnosticVMOptions -XX:+G1SummarizeConcMark
基本上借助普通的GC日志就能发现问题。如果想详细了解每个阶段的耗时、每个Region的回收效率,那么就可打开这些选项。比如统计并发标记各阶段耗时,统计每个Region的数据是如何分布的。以此为基础来调整停顿时间、RegionSize等参数来做优化。
//YGC各阶段耗时
//GC线程还是8个,本次GC造成的停顿时间是0.00178,很明显是在20ms内的;
0.189: [GC pause (G1 Evacuation Pause) (young), 0.0020256 secs]
[Parallel Time: 1.4 ms, GC Workers: 8]
//注意,下面这一行信息是说我们这8个线程开始的时间,min表示最早开始的线程时间,avg表示平均开始时间,max表示最晚开始时间,diff表示最早和最晚的时间差,这个diff越大就说明线程启动时间越不均衡,一般这个地方的时间不会差别很大;
[GC Worker Start (ms): 188.7 188.7 188.7 188.7 188.8 188.8 188.8 188.8
Min: 188.7, Avg: 188.7, Max: 188.8, Diff: 0.1]
//8个线程对根(GC Roots)扫描处理的时间,可以注意到,根的类型非常多,除了经常说的Java静态变量、方法局部变量外还有很多其他的一些根
[Ext Root Scanning (ms): 0.3 1.2 0.2 0.0 0.2 0.2 0.2 0.2
Min: 0.0, Avg: 0.3, Max: 1.2, Diff: 1.2, Sum: 2.5]
[Thread Roots (ms): 0.0 1.2 0.1 0.0 0.0 0.0 0.1 0.0
Min: 0.0, Avg: 0.2, Max: 1.2, Diff: 1.2, Sum: 1.5]
[StringTable Roots (ms): 0.0 0.0 0.1 0.0 0.1 0.1 0.1 0.2
Min: 0.0, Avg: 0.1, Max: 0.2, Diff: 0.2, Sum: 0.6]
[Universe Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[JNI Handles Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[ObjectSynchronizer Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[FlatProfiler Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Management Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[SystemDictionary Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[CLDG Roots (ms): 0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.3, Diff: 0.3, Sum: 0.3]
[JVMTI Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[CodeCache Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[CM RefProcessor Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Wait For Strong CLD (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Weak CLD Roots (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[SATB Filtering (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
//更新RSet的操作,前面提过要先更新RSet后,再进行RSet的扫描,其实在这个日志里就能看的出来;
//对象引用关系更新之后,要通过Refine线程 + 系统程序的线程 + DCQ更新RSet
//下面显示的是GC线程时间都为0.0,说明这块更新没问题
//因为引用更新导致RSet更新的这些操作,都已经由Refine线程提前处理完了;
[Update RS (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Processed Buffers: 0 0 0 0 0 0 0 0
Min: 0, Avg: 0.0, Max: 0, Diff: 0, Sum: 0]
//扫描RSet,处理存活对象;
[Scan RS (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Code Root Scanning (ms): 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
//这一步很关键,对象拷贝,就是把存活对象拷贝到一个新的分区,可以看出来,拷贝消耗的时间非常长;
//注意:对象拷贝消耗的时间占了GC绝大部分的时间
[Object Copy (ms): 0.5 0.0 0.6 0.0 0.7 0.6 0.6 0.6
Min: 0.0, Avg: 0.5, Max: 0.7, Diff: 0.7, Sum: 3.8]
//GC线程关闭的一些信息
[Termination (ms): 0.4 0.0 0.4 0.4 0.4 0.4 0.4 0.4
Min: 0.0, Avg: 0.4, Max: 0.4, Diff: 0.4, Sum: 2.8]
[Termination Attempts: 3 1 12 1 16 9 11 9
Min: 1, Avg: 7.8, Max: 16, Diff: 15, Sum: 62]
//GC线程的一些其他工作的处理时间,通常是JVM析构释放资源
[GC Worker Other (ms): 0.0 0.1 0.1 0.9 0.1 0.1 0.0 0.0
Min: 0.0, Avg: 0.2, Max: 0.9, Diff: 0.9, Sum: 1.3]
//并行GC过程花费的总体时间;
[GC Worker Total (ms): 1.3 1.3 1.3 1.3 1.3 1.3 1.3 1.3
Min: 1.3, Avg: 1.3, Max: 1.3, Diff: 0.1, Sum: 10.4]
[GC Worker End (ms): 190.0 190.0 190.0 190.1 190.0 190.1 190.0 190.1
Min: 190.0, Avg: 190.0, Max: 190.1, Diff: 0.0]
[Code Root Fixup: 0.0 ms]
[Code Root Purge: 0.0 ms]
//清理卡表
[Clear CT: 0.2 ms]
[Other: 0.5 ms]
//选择Cset时间,对YGC来说,一般都是0,因为不需要选
[Choose CSet: 0.0 ms]
[Ref Proc: 0.2 ms]
[Ref Enq: 0.0 ms]
//重构RSet花费的时间
[Redirty Cards: 0.2 ms]
[Parallel Redirty: 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Min: 0.0, Avg: 0.0, Max: 0.0, Diff: 0.0, Sum: 0.0]
[Redirtied Cards: 0 0 0 0 0 0 0 0
Min: 0, Avg: 0.0, Max: 0, Diff: 0, Sum: 0]
//大对象的处理时间
[Humongous Register: 0.0 ms]
[Humongous Total: 0]
[Humongous Candidate: 0]
//如果大对象要回收,回收花费的时间以及回收了多少个大对象
[Humongous Reclaim: 0.0 ms]
[Humongous Reclaimed: 0]
//释放CSet中的分区花费的时间
[Free CSet: 0.0 ms]
[Young Free CSet: 0.0 ms]
[Non-Young Free CSet: 0.0 ms]
//Mixed GC各阶段耗时
//初始标记之后会进入并发标记;
//下面这几行实际上做的是从YHR中的Survivor分区开始进行并发标记的GC Roots扫描;
//可以理解成,找到Survivor区的所有存活对象,并作为根对象,这一部分还没有对GC Roots直接引用的老年代对象做扫描
8.311: [GC concurrent-root-region-scan-start]
//扫描结束,用时0.0000467秒,这个扫描过程需要注意的是:它是和应用程序同时进行的,并且有多个线程参与
8.311: [GC concurrent-root-region-scan-end, 0.0000467 secs]
//并发标记的全量扫描过程,就是以Survivor + GC Roots(方法栈里面的局部变量,静态变量等),标记整个堆里面的所有存活对象
8.311: [GC concurrent-mark-start]
//并发标记结束,花费了0.0018565秒
8.313: [GC concurrent-mark-end, 0.0018556 secs]
//最终标记阶段,标记程序创建出来的新对象,软引用、弱引用等的处理;
8.313: [GC remark 8.313: [Finalize Marking, 0.0001905 secs] 8.313: [GC ref-proc, 0.0003040 secs] 8.313: [Unloading 8.313: [System Dictionary Unloading, 0.0000130 secs] 8.313: [Parallel Unloading, 0.0005763 secs] 8.314: [Deallocate Metadata, 0.0000144 secs], 0.0006416 secs], 0.0014267 secs]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//预清理处理,这个地方清除的是全部都是垃圾的那些Region分区,从这里的日志可以看出来,没有清理掉任何一个Region
8.314: [GC cleanup 75M->75M(128M), 0.0016325 secs]
[Times: user=0.00 sys=0.00, real=0.00 secs]
8.FGC的日志信息及日志详情
(1)参数设置
(2)模拟代码
(3)代码运行相关的GC日志
(4)总结
(1)参数设置
具体参数如下:这里先不打印详情,看看Full GC出现的时候是什么时候。
-XX:InitialHeapSize=36M -XX:MaxHeapSize=36M
-XX:G1HeapRegionSize=4M -XX:MaxGCPauseMillis=20
-XX:+PrintGC -XX:+PrintGCTimeStamps -XX:+UseG1GC
-Xloggc:gc.log
(2)模拟代码
public class FGCDemo12 {
private static final ArrayList list = new ArrayList();
public static void main(String[] args) throws Exception {
while(true) {
//设定一个对象为256K,如果太大,会直接走大对象分配无法模拟YGC
byte[] data = new byte[1024 * 1024];
for (int i = 0; i< 36; i++) {
data = new byte[1024 * 1024];
}
//创建n个对象,加入到一个list中,不回收这些数组
for (int i = 0; i<= 36; i++) {
byte[] data2 = new byte[1024 * 1024];
// 一部分字节数组加入到list中
list.add(data2);
}
Thread.sleep(1000);
}
}
}
(3)代码运行相关的GC日志
一.不加打印详情的时候,GC信息如下:
0.160: [GC pause (G1 Evacuation Pause) (young) 4096K->1808K(32M), 0.0028039 secs]
0.163: [GC pause (G1 Evacuation Pause) (young) 5904K->1976K(32M), 0.0017013 secs]
0.165: [GC pause (G1 Evacuation Pause) (young) 6072K->1997K(32M), 0.0012654 secs]
0.167: [GC pause (G1 Evacuation Pause) (young) 6093K->1819K(32M), 0.0012048 secs]
0.171: [GC pause (G1 Evacuation Pause) (young) 13M->2198K(32M), 0.0016891 secs]
0.174: [GC pause (G1 Evacuation Pause) (young) 14M->1900K(32M), 0.0013103 secs]
0.176: [GC pause (G1 Evacuation Pause) (young) 13M->2992K(32M), 0.0016521 secs]
0.179: [GC pause (G1 Evacuation Pause) (young) 14M->14M(32M), 0.0030006 secs]
0.183: [GC pause (G1 Evacuation Pause) (young)-- 30M->32M(32M), 0.0014989 secs]
0.184: [Full GC (Allocation Failure) 32M->23M(32M), 0.0052335 secs]
0.190: [Full GC (Allocation Failure) 23M->23M(32M), 0.0028101 secs]
0.193: [GC pause (G1 Evacuation Pause) (young) 23M->23M(32M), 0.0008639 secs]
0.194: [GC pause (G1 Evacuation Pause) (young) (initial-mark) 23M->23M(32M), 0.0008385 secs]
0.195: [GC concurrent-root-region-scan-start]
0.195: [GC concurrent-root-region-scan-end, 0.0000051 secs]
0.195: [GC concurrent-mark-start]
//第一次Full GC压缩出1M的空间,从23M到22M;第二次Full GC没法压缩出空间了
0.195: [Full GC (Allocation Failure) 23M->22M(32M), 0.0062830 secs]
0.201: [Full GC (Allocation Failure) 22M->22M(32M), 0.0028600 secs]
//终止并发标记
0.204: [GC concurrent-mark-abort]
0.204: [GC pause (G1 Evacuation Pause) (young) 22M->22M(32M), 0.0006367 secs]
0.205: [GC pause (G1 Evacuation Pause) (young) (initial-mark) 22M->22M(32M), 0.0007129 secs]
0.206: [GC concurrent-root-region-scan-start]
0.206: [GC concurrent-root-region-scan-end, 0.0000050 secs]
0.206: [GC concurrent-mark-start]
//在并发标记开始时,进入了Full GC,并且进行了两次Full GC,最终并发标记终止,抛出了OOM异常;
//说明我们对Full GC的分析是正确的,会分两次GC,第一次不处理软引用,第二次处理;
0.206: [Full GC (Allocation Failure) 22M->22M(32M), 0.0024892 secs]
0.208: [Full GC (Allocation Failure) 22M->22M(32M), 0.0020168 secs]
0.210: [GC concurrent-mark-abort]
二.修改参数为打印日志详情:
-XX:InitialHeapSize=36M -XX:MaxHeapSize=36M
-XX:G1HeapRegionSize=4M -XX:MaxGCPauseMillis=20
-XX:+PrintGC -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+UseG1GC
-Xloggc:gc.log
在添加了打印日志详情 -XX:+PrintGCDetails 参数之后,日志输出如下:
CommandLine flags: -XX:G1HeapRegionSize=4194304 -XX:InitialHeapSize=37748736 -XX:MaxGCPauseMillis=20 -XX:MaxHeapSize=37748736 -XX:+PrintGC -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+UseCompressedClassPointers -XX:+UseCompressedOops -XX:+UseG1GC -XX:-UseLargePagesIndividualAllocation
//第1次YGC
0.164: [GC pause (G1 Evacuation Pause) (young), 0.0015655 secs]
[Parallel Time: 1.1 ms, GC Workers: 8]
...
[Eden: 4096.0K(4096.0K)->0.0B(4096.0K) Survivors: 0.0B->4096.0K Heap: 4096.0K(36.0M)->1896.1K(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//第2次YGC
0.166: [GC pause (G1 Evacuation Pause) (young), 0.0016293 secs]
[Parallel Time: 1.1 ms, GC Workers: 8]
...
[Eden: 4096.0K(4096.0K)->0.0B(4096.0K) Survivors: 4096.0K->4096.0K Heap: 5992.1K(36.0M)->2202.0K(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//第3次YGC
0.168: [GC pause (G1 Evacuation Pause) (young), 0.0014074 secs]
[Parallel Time: 0.9 ms, GC Workers: 8]
...
[Eden: 4096.0K(4096.0K)->0.0B(4096.0K) Survivors: 4096.0K->4096.0K Heap: 6298.0K(36.0M)->1777.8K(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//第4次YGC
0.170: [GC pause (G1 Evacuation Pause) (young), 0.0013150 secs]
[Parallel Time: 0.9 ms, GC Workers: 8]
...
[Eden: 4096.0K(4096.0K)->0.0B(8192.0K) Survivors: 4096.0K->4096.0K Heap: 5873.8K(36.0M)->1948.1K(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//第5次YGC
0.173: [GC pause (G1 Evacuation Pause) (young), 0.0016268 secs]
[Parallel Time: 1.2 ms, GC Workers: 8]
...
[Eden: 8192.0K(8192.0K)->0.0B(16.0M) Survivors: 4096.0K->4096.0K Heap: 10140.1K(36.0M)->2080.7K(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//第6次YGC
0.178: [GC pause (G1 Evacuation Pause) (young), 0.0015187 secs]
[Parallel Time: 1.1 ms, GC Workers: 8]
...
[Eden: 16.0M(16.0M)->0.0B(16.0M) Survivors: 4096.0K->4096.0K Heap: 18.0M(36.0M)->2069.6K(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//第7次YGC
0.181: [GC pause (G1 Evacuation Pause) (young), 0.0018350 secs]
...
[Eden: 16.0M(16.0M)->0.0B(16.0M) Survivors: 4096.0K->4096.0K Heap: 18.0M(36.0M)->6165.5K(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//第8次YGC
0.183: [GC pause (G1 Evacuation Pause) (young) (to-space exhausted), 0.0032817 secs]
[Parallel Time: 2.7 ms, GC Workers: 8]
...
[Eden: 16.0M(16.0M)->0.0B(8192.0K) Survivors: 4096.0K->4096.0K Heap: 22.0M(36.0M)->24.0M(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//第9次YGC,但在这里开启了Mixed GC的初始标记,然后紧接着进入了并发标记阶段
0.187: [GC pause (G1 Evacuation Pause) (young) (initial-mark) (to-space exhausted), 0.0013577 secs]
[Parallel Time: 0.8 ms, GC Workers: 8]
...
[Eden: 8192.0K(8192.0K)->0.0B(8192.0K) Survivors: 4096.0K->0.0B Heap: 32.0M(36.0M)->36.0M(36.0M)]
[Times: user=0.13 sys=0.00, real=0.00 secs]
//进入Mixed GC的并发标记阶段之后,开始并发标记时,直接进入了Full GC状态
0.189: [GC concurrent-root-region-scan-start]
0.189: [GC concurrent-root-region-scan-end, 0.0000047 secs]
0.189: [GC concurrent-mark-start]
//可以看到Full GC第一次进行时,还是回收了一些垃圾的,并且对堆内存进行了整理;
//在Full GC之后,就没有Eden区、Survivor区的具体细节了;
//因为Full GC结束后,会把所有分区标记成Old,然后再重新选择一些Region成为Eden区;
0.189: [Full GC (Allocation Failure) 36M->23M(36M), 0.0051574 secs]
[Eden: 0.0B(8192.0K)->0.0B(4096.0K) Survivors: 0.0B->0.0B Heap: 36.0M(36.0M)->23.7M(36.0M)], [Metaspace: 3445K->3445K(1056768K)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//然后此时并发标记就直接终止,不再进行了;
0.194: [GC concurrent-mark-abort]
//程序运行一段时间后,Eden区又一次填满,此时继续进入YGC
0.194: [GC pause (G1 Evacuation Pause) (young) (to-space exhausted), 0.0007054 secs]
[Parallel Time: 0.2 ms, GC Workers: 8]
...
//注意:在Full GC之后,Eden区还是一个Region,并且此次YGC之后,没有回收掉任何垃圾;
[Eden: 4096.0K(4096.0K)->0.0B(8192.0K) Survivors: 0.0B->0.0B Heap: 27.7M(36.0M)->27.7M(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//接着程序继续运行,只能再此触发第一次的Full GC,并且仅仅压缩整理出了1M的空间
0.195: [Full GC (Allocation Failure) 27M->26M(36M), 0.0026733 secs]
[Eden: 0.0B(8192.0K)->0.0B(4096.0K) Survivors: 0.0B->0.0B Heap: 27.7M(36.0M)->26.7M(36.0M)], [Metaspace: 3445K->3445K(1056768K)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//然后再此触发第二次的Full GC,此次回收会把软引用进行处理;
//而我们的程序没有软引用,所以此次Full GC也无法腾出足够的空间;
0.198: [Full GC (Allocation Failure) 26M->26M(36M), 0.0020266 secs]
[Eden: 0.0B(4096.0K)->0.0B(4096.0K) Survivors: 0.0B->0.0B Heap: 26.7M(36.0M)->26.7M(36.0M)], [Metaspace: 3445K->3445K(1056768K)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//然后再次进入YGC,因为Eden区为0,此时再次尝试分配失败
0.200: [GC pause (G1 Evacuation Pause) (young), 0.0007709 secs]
[Parallel Time: 0.3 ms, GC Workers: 8]
...
[Eden: 0.0B(4096.0K)->0.0B(8192.0K) Survivors: 0.0B->0.0B Heap: 26.7M(36.0M)->26.7M(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//然后再次进入YGC,开启Mixed GC的初始标记
0.201: [GC pause (G1 Evacuation Pause) (young) (initial-mark), 0.0006809 secs]
[Parallel Time: 0.3 ms, GC Workers: 8]
...
[Eden: 0.0B(8192.0K)->0.0B(8192.0K) Survivors: 0.0B->0.0B Heap: 26.7M(36.0M)->26.7M(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
0.202: [GC concurrent-root-region-scan-start]
0.202: [GC concurrent-root-region-scan-end, 0.0000054 secs]
0.202: [GC concurrent-mark-start]
0.202: [Full GC (Allocation Failure) 26M->25M(36M), 0.0067177 secs]
[Eden: 0.0B(8192.0K)->0.0B(4096.0K) Survivors: 0.0B->0.0B Heap: 26.7M(36.0M)->25.7M(36.0M)], [Metaspace: 3445K->3445K(1056768K)]
[Times: user=0.01 sys=0.00, real=0.01 secs]
//触发Full GC
0.209: [Full GC (Allocation Failure) 25M->25M(36M), 0.0022560 secs]
[Eden: 0.0B(4096.0K)->0.0B(4096.0K) Survivors: 0.0B->0.0B Heap: 25.7M(36.0M)->25.7M(36.0M)], [Metaspace: 3445K->3445K(1056768K)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//终止Mixed GC的并发标记
0.211: [GC concurrent-mark-abort]
//触发YGC
0.211: [GC pause (G1 Evacuation Pause) (young), 0.0006518 secs]
[Parallel Time: 0.3 ms, GC Workers: 8]
...
[Eden: 0.0B(4096.0K)->0.0B(4096.0K) Survivors: 0.0B->0.0B Heap: 25.7M(36.0M)->25.7M(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
////然后再次进入YGC,开启Mixed GC的初始标记
0.212: [GC pause (G1 Evacuation Pause) (young) (initial-mark), 0.0006747 secs]
[Parallel Time: 0.3 ms, GC Workers: 8]
...
[Eden: 0.0B(4096.0K)->0.0B(4096.0K) Survivors: 0.0B->0.0B Heap: 25.7M(36.0M)->25.7M(36.0M)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
0.213: [GC concurrent-root-region-scan-start]
0.213: [GC concurrent-root-region-scan-end, 0.0000046 secs]
0.213: [GC concurrent-mark-start]
//触发第一次Full GC
0.213: [Full GC (Allocation Failure) 25M->25M(36M), 0.0025106 secs]
[Eden: 0.0B(4096.0K)->0.0B(4096.0K) Survivors: 0.0B->0.0B Heap: 25.7M(36.0M)->25.7M(36.0M)], [Metaspace: 3445K->3445K(1056768K)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//触发第二次Full GC
0.216: [Full GC (Allocation Failure) 25M->25M(36M), 0.0020506 secs]
[Eden: 0.0B(4096.0K)->0.0B(4096.0K) Survivors: 0.0B->0.0B Heap: 25.7M(36.0M)->25.7M(36.0M)], [Metaspace: 3445K->3445K(1056768K)]
[Times: user=0.00 sys=0.00, real=0.00 secs]
//终止Mixed GC的并发标记
0.218: [GC concurrent-mark-abort]
最终经过多次尝试之后,无法分配,只能OOM。可以看到G1对于OOM其实是比较慎重的。经过很多次YGC并发标记,多次FGC的过程后,才最终抛出OOM异常。所以并不是两次FGC失败之后就进入OOM了。
注意:FGC结束后,会把所有分区标记成Old,然后再重新选择一些Region成为Eden区。
(4)总结
使用G1时发生FGC要比使用ParNew + CMS时发生FGC更加困难,并且尝试的次数日志上来看也非常多。FGC的整个过程比较简单粗暴,如果实在无法压缩出空间,就会OOM。
G1原理—7.G1的GC日志分析解读的更多相关文章
- 【转】gc日志分析工具
性能测试排查定位问题,分析调优过程中,会遇到要分析gc日志,人肉分析gc日志有时比较困难,相关图形化或命令行工具可以有效地帮助辅助分析. Gc日志参数 通过在tomcat启动脚本中添加相关参数生成gc ...
- GC之七--gc日志分析工具
性能测试排查定位问题,分析调优过程中,会遇到要分析gc日志,人肉分析gc日志有时比较困难,相关图形化或命令行工具可以有效地帮助辅助分析. Gc日志参数 通过在tomcat启动脚本中添加相关参数生成gc ...
- JAVA 从GC日志分析堆内存 第七节
JAVA 从GC日志分析堆内存 第七节 在上一章中,我们只设置了整个堆的内存大小.但是我们知道,堆又分为了新生代,年老代.他们之间的内存怎么分配呢?新生代又分为Eden和Survivor,他们的比 ...
- jvm的GC日志分析 [转]
jvm的GC日志分析 标签: jvm内存javagc 2015-06-22 16:37 1566人阅读 评论(1) 收藏 举报 分类: Java(4) JVM的GC日志的主要参数包括如下几个: ...
- GC日志分析详解
点击返回上层目录 原创声明:作者:Arnold.zhao 博客园地址:https://www.cnblogs.com/zh94 GC日志分析详解 以ParallelGC为例,YoungGC日志解释如下 ...
- Android内存优化(二)DVM和ART的GC日志分析
相关文章 Android内存优化系列 Java虚拟机系列 前言 在Java虚拟机(三)垃圾标记算法与Java对象的生命周期这篇文章中,提到了Java虚拟机的GC日志.DVM和ART的GC日志与Java ...
- java虚拟机(十一)--GC日志分析
GC相关:java虚拟机(六)--垃圾收集器和内存分配策略 java虚拟机(五)--垃圾回收机制GC 打印日志相关参数: -XX:+PrintGCDetails -XX:PrintGCTimestam ...
- GC日志分析
JVM的GC日志的主要參数包含例如以下几个: -XX:+PrintGC 输出GC日志 -XX:+PrintGCDetails 输出GC的具体日志 -XX:+PrintGCTimeStamps 输出GC ...
- JVM调优——之CMS GC日志分析
最近在学习JVM和GC调优,今天总结下CMS的一些特点和要点,让我们先简单的看下整个堆年轻代和年老代的垃圾收集器组合(以下配合java8完美支持,其他版本可能稍有不同),其中标红线的则是我们今天要着重 ...
- JVM(7)之 从GC日志分析堆内存
开发十年,就只剩下这套架构体系了! >>> 在前面的文章中,我们只设置了整个堆的内存大小.但是我们知道,堆又分为了新生代,年老代.他们之间的内存怎么分配呢?新生代又分为Eden和 ...
随机推荐
- AnytimeCL:难度加大,支持任意持续学习场景的新方案 | ECCV'24
来源:晓飞的算法工程笔记 公众号,转载请注明出处 论文: Anytime Continual Learning for Open Vocabulary Classification 论文地址:http ...
- (系列十一)Vue3框架中路由守卫及请求拦截(实现前后端交互)
说明 该文章是属于OverallAuth2.0系列文章,每周更新一篇该系列文章(从0到1完成系统开发). 该系统文章,我会尽量说的非常详细,做到不管新手.老手都能看懂. 说明:OverallAuth2 ...
- REST API 已经 25 岁了:它是如何形成的,将来可能会怎样?
原文:https://journal.hexmos.com/rest-turns-25/ 原题:REST APIs Turn 25: How They Came To Be and What Coul ...
- CF1515F Phoenix and Earthquake
CF1515F Phoenix and Earthquake 证明题. 思路 考虑不合法的情况,如果 \(\sum a_i < (n-1)\times x\),肯定是不合法的. 再考虑对于一个可 ...
- 全新向量数据库SQL Server 2025:带你迈入AI驱动的数据未来
全新向量数据库SQL Server 2025:带你迈入AI驱动的数据未来 上次大家下单的<微软憋大招:SQL Server + Copilot = 地表最强AI数据库!> 抱怨迟迟没有发货 ...
- mysql与mongodb、redis的性能对比,包含源代码
最近因为工作需要,所以研究了一下mongodb,但是不知道它的真实性能怎么样,所以对比mysql做了一些测试,现在只是插入操作,更新还没来得及弄,贴出来和大家分享一下(已经补齐redis部分) 1.数 ...
- Java开发
总结java开发中知识点和问题点 基础: 常用加解密算法: [md5] import java.security.MessageDigest; public static final String e ...
- python语言实现_通过端口转发实现跨网络(多网络之间)通信_science_network
本文使用python语言实现了一个端口转发的程序,该程序可以实现多网络之间的信息通信,当然这里有个前提,那就是多个网络都在一台主机上有可以连通的端口. 之所以有这个编写代码的需求,是因为最近使用的sc ...
- ZCMU-1156
思路: 要改变的是一个范围的情况,所以正常情况下会超时. 查阅后知道应该用一个叫做树状数组的结构. 查阅和树状数组的后续情况 这个也不错 注意: 我没怎么看懂,可能没太仔细看. 树状数组当中存在的是前 ...
- Node.js Express 框架(2)
1.读取文件并返回给客户端 res.sendFile(path):读取文件并返回给客户端,适合静态页面 app.get("/",function(req,res){ res.sen ...