简单来说,对象内存分配主要是在堆中分配。但是分配的规则并不是固定的,取决于使用的收集器组合以及JVM内存相关参数的设定
以下介绍几条基本规则(使用的ParNew+Serial Old收集器组合):
一,对象优先在新生代Eden区分配
-
- public class test {
- static int mb = 1024*1024;
-
- public static void main(String[] args) {
- byte[] b1 = new byte[2*mb];
- System.out.println("b1 over");
- byte[] b2 = new byte[2*mb];
- System.out.println("b2 over");
- byte[] b3 = new byte[2*mb];
- System.out.println("b3 over");
- byte[] b4 = new byte[4*mb];
- System.out.println("b4 over");
- }
- }
堆内存大小为20M,不可自动扩展,新生代内存大小为10M,根据默认值,Eden区:Survivor区为8:1,Eden区大小应为:10M*8/10=8129KB,Survivor区大小应为1024KB,新生代总可用内存应为9216KB
当b3分配完成后,新生代将使用6M内存(6144KB,b1+b2+b3),同时申请b4的4M=4096KB内存,此时新生代的可用内存为9216-6144=3072KB,不足以分配b4的空间,则触发一次Minor GC回收新生代内存空间,由于b1、b2以及b3都为存活状态,并且剩余的一个Survivor区无法装下b1、b2和b3,则新生代会租借老年代的区域,并将b1、b2和b3移动至租借区域,然后新生代完成Minor GC。由于此时新生代已经没有对象存放其中,剩余大量内存,则b4将在新生代中分配
- b1 over
- b2 over
- b3 over
- {Heap before GC invocations=0 (full 0):
- par new generation total 9216K, used 6487K [0x03b30000, 0x04530000, 0x04530000)//b1+b2+b3,占6M
- eden space 8192K, 79% used [0x03b30000, 0x04185f60, 0x04330000)
- from space 1024K, 0% used [0x04330000, 0x04330000, 0x04430000)
- to space 1024K, 0% used [0x04430000, 0x04430000, 0x04530000)
- tenured generation total 10240K, used 0K [0x04530000, 0x04f30000, 0x04f30000)//老年代为空
- the space 10240K, 0% used [0x04530000, 0x04530000, 0x04530200, 0x04f30000)
- compacting perm gen total 12288K, used 2105K [0x04f30000, 0x05b30000, 0x08f30000)
- the space 12288K, 17% used [0x04f30000, 0x0513e478, 0x0513e600, 0x05b30000)
- No shared spaces configured.
- [GC [ParNew: 6487K->150K(9216K), 0.0092952 secs] 6487K->6294K(19456K), 0.0093314 secs] [Times: user=0.00 sys=0.00, real=0.00 secs]//对象仍处于存活状态,新生代无足够的空间完成Minor GC,只能租借老年代的空间,将b1、b2和b3移动至老年代
- Heap after GC invocations=1 (full 0):
- par new generation total 9216K, used 150K [0x03b30000, 0x04530000, 0x04530000)//新生代几乎被清空
- eden space 8192K, 0% used [0x03b30000, 0x03b30000, 0x04330000)
- from space 1024K, 14% used [0x04430000, 0x04455a10, 0x04530000)
- to space 1024K, 0% used [0x04330000, 0x04330000, 0x04430000)
- tenured generation total 10240K, used 6144K [0x04530000, 0x04f30000, 0x04f30000)//b1+b2+b3
- the space 10240K, 60% used [0x04530000, 0x04b30030, 0x04b30200, 0x04f30000)
- compacting perm gen total 12288K, used 2105K [0x04f30000, 0x05b30000, 0x08f30000)
- the space 12288K, 17% used [0x04f30000, 0x0513e478, 0x0513e600, 0x05b30000)
- No shared spaces configured.
- }
- b4 over
- Heap
- par new generation total 9216K, used 4410K [0x03b30000, 0x04530000, 0x04530000)//b4
- eden space 8192K, 54% used [0x03b30000, 0x03f82008, 0x04330000)
- from space 1024K, 14% used [0x04430000, 0x04455a10, 0x04530000)
- to space 1024K, 0% used [0x04330000, 0x04330000, 0x04430000)
- tenured generation total 10240K, used 6144K [0x04530000, 0x04f30000, 0x04f30000)//b1+b2+b3
- the space 10240K, 60% used [0x04530000, 0x04b30030, 0x04b30200, 0x04f30000)
- compacting perm gen total 12288K, used 2116K [0x04f30000, 0x05b30000, 0x08f30000)
- the space 12288K, 17% used [0x04f30000, 0x051413c8, 0x05141400, 0x05b30000)
- No shared spaces configured.
二,大对象直接进入老年代
为了避免内存回收时大对象在Eden区和2个Survivor区之间的拷贝(ParNew收集器使用复制算法),同时为了避免为了提供足够的内存空间而提前触发的GC,虚拟机提供了-XX:PretenureSizeThreshold(该设置只对Serial和ParNew收集器生效)参数,大于该参数设置值的对象将直接在老年代分配
-
-
- public class test {
- static int mb = 1024*1024;
-
- public static void main(String[] args) {
- byte[] b1 = new byte[3*mb];
- System.out.println("b1 over");
- }
- }
由于设置超过2M(2*1024*1024=2097152B)的对象直接在老年代分配,故b1将分配在老年代上
- b1 over
- Heap
- par new generation total 9216K, used 507K [0x03b50000, 0x04550000, 0x04550000)//新生代几乎为空
- eden space 8192K, 6% used [0x03b50000, 0x03bcef00, 0x04350000)
- from space 1024K, 0% used [0x04350000, 0x04350000, 0x04450000)
- to space 1024K, 0% used [0x04450000, 0x04450000, 0x04550000)
- tenured generation total 10240K, used 3072K [0x04550000, 0x04f50000, 0x04f50000)//老年代使用了3*1024K内存
- the space 10240K, 30% used [0x04550000, 0x04850010, 0x04850200, 0x04f50000)
- compacting perm gen total 12288K, used 2110K [0x04f50000, 0x05b50000, 0x08f50000)
- the space 12288K, 17% used [0x04f50000, 0x0515f8c8, 0x0515fa00, 0x05b50000)
- No shared spaces configured.
三,长期存活对象将进入老年代
由于虚拟机垃圾收集是基于“分代算法”的,故虚拟机必须能够识别哪些对象存放在新生代,哪些对象应该存放在老年代
虚拟机设计了一个对象年龄计数器,如果对象在Eden区出生并且经过第一次Minor GC后依然存活,并且可以被Survivor区容纳,就会被复制至Survivor区并将对象年龄设置为1。以后对象每熬过一次Minor GC,对象年龄便+1。当对象年龄超过对象晋升老年代的年龄阀值(该阀值默认为15)时,便会晋升至老年代,何时晋升,我们接下来研究
虚拟机提供了-XX:MaxTenuringThreshold参数设置晋升阀值
-
-
- public class test {
- static int mb = 1024*1024;
-
- public static void main(String[] args) {
- System.out.println("step 1");
- byte[] b1 = new byte[1*mb/4];
- System.out.println("step 2");
- byte[] b2 = new byte[4*mb];
- System.out.println("step 3");
- byte[] b3 = new byte[4*mb];
- System.out.println("step 4");
- b3 = null;
- System.out.println("step 5");
- b3 = new byte[4*mb];
- }
- }
b1、b2正常分配。在step3,新生代将没有足够的内存分配b3所需的4M空间,故引发一次Minor GC。b1只有256KB,可以放置在Survivor区中,故复制b1到Survivor区中,b2为4M,无法放置到Survivor区中,故租借老年代4M内存放置b2,回收新生代内存空间,b1经历了一次Minor GC后依然存活,故年龄变为1。
在step4,分配给b3对象的内存空间依然被占用,只是将b3对象的引用置为空,由于不涉及到内存分配,故而不涉及到GC,因此对象的年龄也不会发生变化
在step5,重新给b3对象分配4M空间,由于新生代没有足够内存,故引发Minor GC,step3分配给b3的4M内存空间由于不再与存活对象相关联,将被回收,同时,由于b1的年龄到达对象晋升老年代的年龄设置,b1将被移动至老年代
- step 1
- step 2
- step 3
- {Heap before GC invocations=0 (full 0):
- par new generation total 9216K, used 4695K [0x03b80000, 0x04580000, 0x04580000)//b1+b2
- eden space 8192K, 57% used [0x03b80000, 0x04015f50, 0x04380000)
- from space 1024K, 0% used [0x04380000, 0x04380000, 0x04480000)
- to space 1024K, 0% used [0x04480000, 0x04480000, 0x04580000)
- tenured generation total 10240K, used 0K [0x04580000, 0x04f80000, 0x04f80000)//此时老年代为空
- the space 10240K, 0% used [0x04580000, 0x04580000, 0x04580200, 0x04f80000)
- compacting perm gen total 12288K, used 2105K [0x04f80000, 0x05b80000, 0x08f80000)
- the space 12288K, 17% used [0x04f80000, 0x0518e450, 0x0518e600, 0x05b80000)
- No shared spaces configured.
- [GC [ParNew: 4695K->409K(9216K), 0.0049519 secs] 4695K->4505K(19456K), 0.0049944 secs] [Times: user=0.00 sys=0.00, real=0.00 secs]
- Heap after GC invocations=1 (full 0):
- par new generation total 9216K, used 409K [0x03b80000, 0x04580000, 0x04580000)//b1
- eden space 8192K, 0% used [0x03b80000, 0x03b80000, 0x04380000)
- from space 1024K, 39% used [0x04480000, 0x044e6610, 0x04580000)
- to space 1024K, 0% used [0x04380000, 0x04380000, 0x04480000)
- tenured generation total 10240K, used 4096K [0x04580000, 0x04f80000, 0x04f80000)//b2
- the space 10240K, 40% used [0x04580000, 0x04980010, 0x04980200, 0x04f80000)
- compacting perm gen total 12288K, used 2105K [0x04f80000, 0x05b80000, 0x08f80000)
- the space 12288K, 17% used [0x04f80000, 0x0518e450, 0x0518e600, 0x05b80000)
- No shared spaces configured.
- }
- step 4
- step 5
- {Heap before GC invocations=1 (full 0):
- par new generation total 9216K, used 4669K [0x03b80000, 0x04580000, 0x04580000)//b1+b3(step3)
- eden space 8192K, 52% used [0x03b80000, 0x03fa9098, 0x04380000)
- from space 1024K, 39% used [0x04480000, 0x044e6610, 0x04580000)
- to space 1024K, 0% used [0x04380000, 0x04380000, 0x04480000)
- tenured generation total 10240K, used 4096K [0x04580000, 0x04f80000, 0x04f80000)//b2
- the space 10240K, 40% used [0x04580000, 0x04980010, 0x04980200, 0x04f80000)
- compacting perm gen total 12288K, used 2111K [0x04f80000, 0x05b80000, 0x08f80000)
- the space 12288K, 17% used [0x04f80000, 0x0518fe08, 0x05190000, 0x05b80000)
- No shared spaces configured.
- [GC [ParNew: 4669K->43K(9216K), 0.0008256 secs] 8765K->4548K(19456K), 0.0008701 secs] [Times: user=0.00 sys=0.00, real=0.00 secs]
- Heap after GC invocations=2 (full 0):
- par new generation total 9216K, used 43K [0x03b80000, 0x04580000, 0x04580000)//step3分配的b3对象空间被回收
- eden space 8192K, 0% used [0x03b80000, 0x03b80000, 0x04380000)
- from space 1024K, 4% used [0x04380000, 0x0438ad90, 0x04480000)
- to space 1024K, 0% used [0x04480000, 0x04480000, 0x04580000)
- tenured generation total 10240K, used 4505K [0x04580000, 0x04f80000, 0x04f80000)//b1+b2
- the space 10240K, 43% used [0x04580000, 0x049e6590, 0x049e6600, 0x04f80000)
- compacting perm gen total 12288K, used 2111K [0x04f80000, 0x05b80000, 0x08f80000)
- the space 12288K, 17% used [0x04f80000, 0x0518fe08, 0x05190000, 0x05b80000)
- No shared spaces configured.
- }
- Heap
- par new generation total 9216K, used 4303K [0x03b80000, 0x04580000, 0x04580000)//b3(step5)
- eden space 8192K, 52% used [0x03b80000, 0x03fa8fe0, 0x04380000)
- from space 1024K, 4% used [0x04380000, 0x0438ad90, 0x04480000)
- to space 1024K, 0% used [0x04480000, 0x04480000, 0x04580000)
- tenured generation total 10240K, used 4505K [0x04580000, 0x04f80000, 0x04f80000)//b1+b2
- the space 10240K, 43% used [0x04580000, 0x049e6590, 0x049e6600, 0x04f80000)
- compacting perm gen total 12288K, used 2116K [0x04f80000, 0x05b80000, 0x08f80000)
- the space 12288K, 17% used [0x04f80000, 0x051913c8, 0x05191400, 0x05b80000)
- No shared spaces configured.
如果修改MaxTenuringThreshold的值为2,从打印日志中可以发现,最终老年代的内存使用量为4096KB=4M,也就是说b1没有晋升至老年代
上面是Minor GC的运行状况,如果是Full GC呢:
-
-
- public class test {
- static int mb = 1024*1024;
-
- public static void main(String[] args) {
- byte[] b1 = new byte[1*mb/4];
- System.gc();
- }
- }
这里我们使用的是Full GC,也就是老年代的GC。
Full GC通常至少伴随着一次Minor GC(并非绝对),看下面日志,这里的Minor GC应该至少发生了2次,一次Minor GC是不会把b1移动至老年代的
- {Heap before GC invocations=0 (full 0):
- par new generation total 9216K, used 599K [0x03b80000, 0x04580000, 0x04580000)//b1
- eden space 8192K, 7% used [0x03b80000, 0x03c15f40, 0x04380000)
- from space 1024K, 0% used [0x04380000, 0x04380000, 0x04480000)
- to space 1024K, 0% used [0x04480000, 0x04480000, 0x04580000)
- tenured generation total 10240K, used 0K [0x04580000, 0x04f80000, 0x04f80000)//老年代为空
- the space 10240K, 0% used [0x04580000, 0x04580000, 0x04580200, 0x04f80000)
- compacting perm gen total 12288K, used 2104K [0x04f80000, 0x05b80000, 0x08f80000)
- the space 12288K, 17% used [0x04f80000, 0x0518e278, 0x0518e400, 0x05b80000)
- No shared spaces configured.
- [Full GC (System) [Tenured: 0K->404K(10240K), 0.0069434 secs] 599K->404K(19456K), [Perm : 2104K->2104K(12288K)], 0.0069992 secs] [Times: user=0.00 sys=0.00, real=0.00 secs]
- Heap after GC invocations=1 (full 1):
- par new generation total 9216K, used 0K [0x03b80000, 0x04580000, 0x04580000)//新生代为空
- eden space 8192K, 0% used [0x03b80000, 0x03b80000, 0x04380000)
- from space 1024K, 0% used [0x04380000, 0x04380000, 0x04480000)
- to space 1024K, 0% used [0x04480000, 0x04480000, 0x04580000)
- tenured generation total 10240K, used 404K [0x04580000, 0x04f80000, 0x04f80000)//b1
- the space 10240K, 3% used [0x04580000, 0x045e5130, 0x045e5200, 0x04f80000)
- compacting perm gen total 12288K, used 2104K [0x04f80000, 0x05b80000, 0x08f80000)
- the space 12288K, 17% used [0x04f80000, 0x0518e278, 0x0518e400, 0x05b80000)
- No shared spaces configured.
- }
- Heap
- par new generation total 9216K, used 327K [0x03b80000, 0x04580000, 0x04580000)
- eden space 8192K, 4% used [0x03b80000, 0x03bd1f98, 0x04380000)
- from space 1024K, 0% used [0x04380000, 0x04380000, 0x04480000)
- to space 1024K, 0% used [0x04480000, 0x04480000, 0x04580000)
- tenured generation total 10240K, used 404K [0x04580000, 0x04f80000, 0x04f80000)
- the space 10240K, 3% used [0x04580000, 0x045e5130, 0x045e5200, 0x04f80000)
- compacting perm gen total 12288K, used 2116K [0x04f80000, 0x05b80000, 0x08f80000)
- the space 12288K, 17% used [0x04f80000, 0x05191190, 0x05191200, 0x05b80000)
- No shared spaces configured.
四:动态对象年龄判定
为了使内存分配更加灵活,虚拟机并不要求对象年龄达到MaxTenuringThreshold才晋升老年代
如果Survivor区中相同年龄所有对象大小的总和大于Survivor区空间的一半,年龄大于或等于该年龄的对象在Minor GC时将复制至老年代
-
-
- public class Test {
- static int mb = 1024*1024;
-
- public static void main(String[] args) {
- System.out.println("step 1");
- byte[] b1 = new byte[1*mb/4];
- byte[] b3 = new byte[4*mb];
- byte[] b4 = new byte[4*mb];
- System.out.println("step 2");
- byte[] b2 = new byte[1*mb/4];
- b4 = null;
- System.out.println("step 3");
- b4 = new byte[4*mb];
- System.out.println("step 4");
- b4 = null;
- b4 = new byte[4*mb];
- }
- }
先来介绍一个设置-XX:+PrintTenuringDistribution,这个参数很有意思,会在Minor GC时打印Survivor区内存容量的一半,晋升老年代年龄阀值,Survivor区中的对象大小以及对象年龄
根据启动参数的设置,Survivor大小的一半是524288B,也就是512KB。第一次GC后,b1依然存活,故年龄变为1。第二次GC后,b1和b2依然存活,故b1的年龄变为2,b2的年龄为1。b1+b2的大小加起来超过了Survivor区容量的一半,此时会修改Survivor区晋升老年代年龄阀值为2(如果移动年龄为2的对象可以使Survivor去的内存使用降至512KB以内,则只移动年龄为2的对象,否则将会同时移动年龄为1的对象)。第三次GC时,将年龄等于晋升阀值的对象移动至老年代,执行GC,GC结束后,b1依然在Survivor区(当然可能从Survivor from区拷贝至了Survivor to区),此时b1的年龄变为2。这时Survivor区的使用内存没有达到512M,修改Survivor区晋升老年代年龄阀值为参数设置的10。
- step 1
-
- Desired survivor size 524288 bytes, new threshold 10 (max 10)
- - age 1: 412800 bytes, 412800 total
- step 2
- step 3
-
- Desired survivor size 524288 bytes, new threshold 2 (max 10)
- - age 1: 262160 bytes, 262160 total
- - age 2: 412800 bytes, 674960 total
- step 4
-
- Desired survivor size 524288 bytes, new threshold 10 (max 10)
- - age 1: 136 bytes, 136 total
- - age 2: 262160 bytes, 262296 total
最后,为什么在第三次GC后,Survivor区还存在一个大小为136B,年龄为1的被使用内存空间?
我猜测,虽然Minor GC时Survivor区没有足够的空间完成GC时会租借老年代的内存,但是在Survivor区依然保存了一个指向老年代租借内存起始地址的引用
五:空间分配担保
这个前面已经出现过多次了,由于新生代使用复制算法,当Minor GC时如果存活对象过多,无法完全放入Survivor区,就会向老年代借用内存存放对象,以完成Minor GC
在触发Minor GC时,虚拟机会先检测之前GC时租借的老年代内存的平均大小是否大于老年代的剩余内存,如果大于,则将Minor GC变为一次Full GC,如果小于,则查看虚拟机是否允许担保失败(-XX:+/-HandlePromotionFailure。从jdk6.0开始,允许担保失败已变为HotSpot虚拟机所有收集器默认设置,虚拟机将不再识别该参数设置,详见JDK-6990095 : Deprecate and eliminate -XX:-HandlePromotionFailure),如果允许担保失败,则只执行一次Minor GC,否则也要将Minor GC变为一次Full GC(直到GC结束时才能确定到底有多少对象需要被移动至老年代,所以在GC前,只能使用粗略的平均值进行判断)