java - 偶尔遇到长时间的垃圾收集延迟，为什么？

Question

我在处理 Java 垃圾回收问题和解释日志时遇到了困难。

我的应用程序要求 GC 的时间不能超过 2 秒，最好少于 100 毫秒。

根据之前的一些建议，我正在尝试以下命令行选项:

 java -XX:MaxGCPauseMillis=100 -XX:NewRatio=9 -XX:+UseConcMarkSweepGC  -XX:+CMSIncrementalMode -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -server -Xmx9g -Xms9g 

应用有大量的长期存储对象，保存在一个ConcurrentLinkedHashMap中。我偶尔会看到长时间停顿，在最坏的情况下长达 10 秒(这是下面 GC 日志的倒数第二个)!

这是我得到的一些输出:

16938.968: [GC 16938.968: [ParNew: 153343K->17022K(153344K), 7.8608580 secs] 6184328K->6122510K(9420160K) icms_dc=7 , 7.8614100 secs] [Times: user=0.63 sys=0.01, real=7.86 secs] 
16947.087: [GC 16947.087: [ParNew: 153342K->17022K(153344K), 7.2604030 secs] 6258830K->6198642K(9420160K) icms_dc=7 , 7.2609780 secs] [Times: user=0.44 sys=0.00, real=7.27 secs] 
16954.614: [GC 16954.614: [ParNew: 153342K->17024K(153344K), 8.4307620 secs] 6334962K->6274625K(9420160K) icms_dc=7 , 8.4313150 secs] [Times: user=0.62 sys=0.01, real=8.43 secs] 
16963.310: [GC 16963.310: [ParNew: 153344K->17023K(153344K), 6.2588760 secs] 6410945K->6350748K(9420160K) icms_dc=7 , 6.2594290 secs] [Times: user=0.48 sys=0.01, real=6.25 secs] 
16969.834: [GC 16969.834: [ParNew: 153343K->17022K(153344K), 6.0274280 secs] 6487068K->6425868K(9420160K) icms_dc=7 , 6.0279830 secs] [Times: user=0.50 sys=0.01, real=6.03 secs] 
16976.122: [GC 16976.123: [ParNew: 153342K->17022K(153344K), 11.7774620 secs] 6562188K->6503030K(9420160K) icms_dc=7 , 11.7780180 secs] [Times: user=0.43 sys=0.04, real=11.78 secs] 
16988.164: [GC 16988.164: [ParNew: 153342K->17024K(153344K), 10.9477920 secs] 6639350K->6579928K(9420160K) icms_dc=7 , 10.9483440 secs] [Times: user=0.37 sys=0.02, real=10.95 secs] 
16999.371: [GC 16999.372: [ParNew: 153344K->17023K(153344K), 9.8828360 secs] 6716248K->6655886K(9420160K) icms_dc=7 , 9.8833940 secs] [Times: user=0.42 sys=0.01, real=9.88 secs] 
17009.509: [GC 17009.509: [ParNew: 153343K->17023K(153344K), 5.0699960 secs] 6792206K->6727987K(9420160K) icms_dc=7 , 5.0705660 secs] [Times: user=0.55 sys=0.01, real=5.07 secs] 
17014.838: [GC 17014.838: [ParNew: 153343K->17023K(153344K), 6.6411750 secs] 6864307K->6790974K(9420160K) icms_dc=7 , 6.6417400 secs] [Times: user=0.37 sys=0.01, real=6.63 secs] 
17021.735: [GC 17021.735: [ParNew: 153343K->17024K(153344K), 8.0545970 secs] 6927294K->6856409K(9420160K) icms_dc=7 , 8.0551790 secs] [Times: user=0.34 sys=0.03, real=8.05 secs] 
17030.052: [GC 17030.053: [ParNew: 153344K->17023K(153344K), 7.9756730 secs] 6992729K->6922569K(9420160K) icms_dc=7 , 7.9762530 secs] [Times: user=0.34 sys=0.01, real=7.98 secs] 
17038.398: [GC 17038.398: [ParNew: 153343K->17022K(153344K), 12.9613300 secs] 7058889K->6990725K(9420160K) icms_dc=7 , 12.9618850 secs] [Times: user=0.39 sys=0.01, real=12.96 secs] 
17051.630: [GC 17051.630: [ParNew: 153342K->17022K(153344K), 6.8942910 secs] 7127045K->7059607K(9420160K) icms_dc=7 , 6.8948380 secs] [Times: user=0.56 sys=0.02, real=6.89 secs] 
17058.798: [GC 17058.798: [ParNew: 153342K->17024K(153344K), 10.0262190 secs] 7195927K->7126351K(9420160K) icms_dc=7 , 10.0267860 secs] [Times: user=0.37 sys=0.01, real=10.02 secs] 
17069.096: [GC 17069.096: [ParNew: 153344K->17023K(153344K), 10.0419500 secs] 7262671K->7195002K(9420160K) icms_dc=7 , 10.0425020 secs] [Times: user=0.40 sys=0.02, real=10.04 secs] 
17079.410: [GC 17079.410: [ParNew: 153343K->17022K(153344K), 13.5389040 secs] 7331322K->7264275K(9420160K) icms_dc=7 , 13.5394610 secs] [Times: user=0.30 sys=0.01, real=13.54 secs] 
17093.223: [GC 17093.224: [ParNew: 153342K->17023K(153344K), 10.5909450 secs] 7400595K->7330446K(9420160K) icms_dc=7 , 10.5915060 secs] [Times: user=0.33 sys=0.00, real=10.58 secs] 
17104.083: [GC 17104.084: [ParNew: 153343K->17024K(153344K), 5.8420210 secs] 7466766K->7392173K(9420160K) icms_dc=7 , 5.8425920 secs] [Times: user=0.57 sys=0.00, real=5.84 secs] 

我花了几个小时浏览描述 Java GC 调优的各种网页，但没有一个真正让我能够解释上面的日志并提出一个行动方案。非常感谢任何基于我提供的日志的具体建议。

更新:根据以下问题:

机器有 16G 内存，这是来自顶部的信息: 内存:总计 15483904k，已使用 15280084k，空闲 203820k，缓冲区 155684k 交换:总共 2031608k，已使用 1347240k，免费 684368k，缓存 3304044k

这是一次不同的运行，但这是该过程的当前最高输出:

  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
  1016 sanity   17   0 10.2g 6.5g 9464 S    1 44.2  10:24.32 java                      

更新 2: 一些更详细的日志记录，这看起来导致了 400 毫秒的延迟:

{Heap before GC invocations=1331 (full 1):
 par new generation   total 153344K, used 153343K [0x00002aaaae200000, 0x00002aaab8860000, 0x00002aaab8860000)
  eden space 136320K, 100% used [0x00002aaaae200000, 0x00002aaab6720000, 0x00002aaab6720000)
  from space 17024K,  99% used [0x00002aaab77c0000, 0x00002aaab885fff0, 0x00002aaab8860000)
  to   space 17024K,   0% used [0x00002aaab6720000, 0x00002aaab6720000, 0x00002aaab77c0000)
 concurrent mark-sweep generation total 7169664K, used 4258496K [0x00002aaab8860000, 0x00002aac6e200000, 0x00002aac6e200000)
 concurrent-mark-sweep perm gen total 21248K, used 13269K [0x00002aac6e200000, 0x00002aac6f6c0000, 0x00002aac73600000)
484.738: [GC 484.738: [ParNew: 153343K->17022K(153344K), 0.3950480 secs] 4411840K->4341689K(7323008K), 0.3954820 secs] [Times: user=0.49 sys=0.07, real=0.40 secs] 
Heap after GC invocations=1332 (full 1):
 par new generation   total 153344K, used 17022K [0x00002aaaae200000, 0x00002aaab8860000, 0x00002aaab8860000)
  eden space 136320K,   0% used [0x00002aaaae200000, 0x00002aaaae200000, 0x00002aaab6720000)
  from space 17024K,  99% used [0x00002aaab6720000, 0x00002aaab77bfb68, 0x00002aaab77c0000)
  to   space 17024K,   0% used [0x00002aaab77c0000, 0x00002aaab77c0000, 0x00002aaab8860000)
 concurrent mark-sweep generation total 7169664K, used 4324666K [0x00002aaab8860000, 0x00002aac6e200000, 0x00002aac6e200000)
 concurrent-mark-sweep perm gen total 21248K, used 13269K [0x00002aac6e200000, 0x00002aac6f6c0000, 0x00002aac73600000)
}

Answer 1

原来问题是堆被换出到磁盘，延迟是 Java 的 GC 必须等待它被换回。

(大部分)通过将 Linux 的“swappiness”参数设置为 0 来解决。