当流量较高时，为什么vmop（Java GC Options）会增加

Question

我有一个Java的在线服务。如果qps约为2000，则延迟是正常的，但如果qps增加到3000，则延迟会更高。

有没有关于改进的建议，尤其是STW时间？

代码逻辑很简单（作为访问服务）：

1. 解码Tcp请求和Protobuf（每个请求大约200KB）
2. 拆分请求并重新包装protobuf
3. 路由到后向服务并异步等待响应
4. 处理来自后向服务的响应，使用另一个protobuf包装并响应客户端。

期金佛山：

-Djava.net.preferIPv4Stack=true -Xmx6g -Xms6g -Xmn4g 
-Djava.ext.dirs=./lib:/data/home/soft/jdk1.8.0_65/jre/lib/ext 
-Dlog4j.configuration=./conf/log4j.properties -Dio.netty.noPreferDirect=false 
-XX:MaxDirectMemorySize=4g -XX:+UseConcMarkSweepGC -XX:ParallelCMSThreads=8 
-XX:+CMSScavengeBeforeRemark 
-XX:+PrintSafepointStatistics -XX:PrintSafepointStatisticsCount=1 
-XX:OnOutOfMemoryError=./bin/stop.sh -XX:+HeapDumpOnOutOfMemoryError 
-XX:HeapDumpPath=. -XX:+PrintGCApplicationStoppedTime 
-XX:+PrintGCDetails -XX:+PrintGCDateStamps -XX:+PrintGCDetails 
-Xloggc:logs/gc.log -XX:+CMSParallelRemarkEnabled 
-XX:ParallelGCThreads=32

我发现更高的STW是由vmop引起的。测试期间没有发生GC。

998.860: GenCollectForAllocation          [     253          2              2    ]      [     0     0     0     0     6    ]  0   
         vmop                    [threads: total initially_running wait_to_block]    [time: spin block sync cleanup vmop] page_trap_count
999.578: GenCollectForAllocation          [     253          2              2    ]      [     0     0     0     0    16    ]  0   
         vmop                    [threads: total initially_running wait_to_block]    [time: spin block sync cleanup vmop] page_trap_count
1000.273: GenCollectForAllocation          [     253          3              3    ]      [     0     0     0     0    11    ]  0   
         vmop                    [threads: total initially_running wait_to_block]    [time: spin block sync cleanup vmop] page_trap_count
1000.970: GenCollectForAllocation          [     253          2              2    ]      [     0     0     0     0     8    ]  0   
         vmop                    [threads: total initially_running wait_to_block]    [time: spin block sync cleanup vmop] page_trap_count
1001.666: GenCollectForAllocation          [     253          2              2    ]      [     0     0     0     0     7    ]  0   

• 当qps约为2000时，vmop通常低于10ms。 Total time for which application threads were stopped低于0.010秒。
• Qps约为3000，vmop高于10ms，

我的问题是：

1. 是否有关于性能改进的建议，尤其是STW（vmop）时间？
2. 是什么导致高vmop？达到安全点或复制记忆？

更多测试数据：后向服务保持qps = 3000，我启动一两个Java服务进行比较并获得性能差距。落后服务的延迟约为6.8毫秒。一个Java进程的GC日志（每个进程的qps = 3000）：

2019-05-16T14:54:56.742+0800: 67083.398: Total time for which application threads were stopped: 0.0104372 seconds, Stopping threads took: 0.0001939 seconds
2019-05-16T14:54:57.788+0800: 67084.444: [GC (Allocation Failure) 2019-05-16T14:54:57.788+0800: 67084.444: [ParNew: 5054250K->22742K(5662336K), 0.0109608 secs] 5776050K->744839K(7759488K), 0.0111268 secs] [Times: user=0.33 sys=0.01, real=0.02 secs] 
2019-05-16T14:54:57.799+0800: 67084.455: Total time for which application threads were stopped: 0.0122376 seconds, Stopping threads took: 0.0001943 seconds
2019-05-16T14:54:58.848+0800: 67085.504: [GC (Allocation Failure) 2019-05-16T14:54:58.848+0800: 67085.504: [ParNew: 5055958K->26545K(5662336K), 0.0108330 secs] 5778055K->748938K(7759488K), 0.0109914 secs] [Times: user=0.34 sys=0.00, real=0.02 secs] 
2019-05-16T14:54:58.859+0800: 67085.515: Total time for which application threads were stopped: 0.0121166 seconds, Stopping threads took: 0.0002086 seconds
2019-05-16T14:54:59.905+0800: 67086.561: [GC (Allocation Failure) 2019-05-16T14:54:59.905+0800: 67086.561: [ParNew: 5059761K->25093K(5662336K), 0.0115089 secs] 5782154K->747784K(7759488K), 0.0116708 secs] [Times: user=0.36 sys=0.00, real=0.01 secs] 
2019-05-16T14:54:59.917+0800: 67086.573: Total time for which application threads were stopped: 0.0127589 seconds, Stopping threads took: 0.0001810 seconds
2019-05-16T14:55:00.957+0800: 67087.613: [GC (Allocation Failure) 2019-05-16T14:55:00.957+0800: 67087.613: [ParNew: 5058309K->27559K(5662336K), 0.0043041 secs] 5781000K->750551K(7759488K), 0.0044697 secs] [Times: user=0.11 sys=0.00, real=0.01 secs] 
2019-05-16T14:55:00.962+0800: 67087.618: Total time for which application threads were stopped: 0.0055974 seconds, Stopping threads took: 0.0001974 seconds
2019-05-16T14:55:02.004+0800: 67088.660: [GC (Allocation Failure) 2019-05-16T14:55:02.004+0800: 67088.660: [ParNew: 5060775K->18771K(5662336K), 0.0032344 secs] 5783767K->742062K(7759488K), 0.0033973 secs] [Times: user=0.08 sys=0.00, real=0.00 secs] 

两个Java进程的GC日志（每个进程的qps = 1500）：

2019-05-16T14:49:19.774+0800: 66746.430: Total time for which application threads were stopped: 0.0075090 seconds, Stopping threads took: 0.0001816 seconds
2019-05-16T14:49:21.879+0800: 66748.535: [GC (Allocation Failure) 2019-05-16T14:49:21.879+0800: 66748.535: [ParNew: 5047304K->15661K(5662336K), 0.0070949 secs] 5601138K->569792K(7759488K), 0.0072536 secs] [Times: user=0.19 sys=0.00, real=0.00 secs] 
2019-05-16T14:49:21.886+0800: 66748.542: Total time for which application threads were stopped: 0.0084102 seconds, Stopping threads took: 0.0001858 seconds
2019-05-16T14:49:23.977+0800: 66750.633: [GC (Allocation Failure) 2019-05-16T14:49:23.977+0800: 66750.633: [ParNew: 5048877K->10672K(5662336K), 0.0033657 secs] 5603008K->565100K(7759488K), 0.0035116 secs] [Times: user=0.08 sys=0.00, real=0.01 secs] 
2019-05-16T14:49:23.981+0800: 66750.637: Total time for which application threads were stopped: 0.0045580 seconds, Stopping threads took: 0.0001763 seconds
2019-05-16T14:49:26.090+0800: 66752.746: [GC (Allocation Failure) 2019-05-16T14:49:26.090+0800: 66752.746: [ParNew2019-05-16T14:49:23.981+0800: 66750.637: Total time for which application threads were stopped: 0.0045580 seconds, Stopping threads took: 0.0001763 seconds
2019-05-16T14:49:26.090+0800: 66752.746: [GC (Allocation Failure) 2019-05-16T14:49:26.090+0800: 66752.746: [ParNew: 5043888K->9061K(5662336K), 0.0036848 secs] 5598316K->563786K(7759488K), 0.0038302 secs] [Times: user=0.08 sys=0.00, real=0.00 secs] 

Answer 1

你的因果关系倒退，它不是由vmop引起的。 vmop是垃圾收集器请求的安全点，在该安全点下它可以执行其停止世界任务。

如果要分析GC行为，则需要先查看GC日志，然后再查看安全点。