海量连接服务端CMS调优记

应用:shark-美团点评移动端长连代理,每日接受移动端请求约150亿

应用特点:

  • QPS比较高,新生代增长飞快

  • 用户的连接需要维持一段时间

  • 单机需要维持海量连接,大概在几十万的级别

以上三个特点导致有大量小对象聚集在old区,高峰期old区域增长非常快,对象在一段时间内必然消亡

初始的线上gc的情况如下(备注:相关截图来自点评的APM软件CAT: https://github.com/dianping/cat):

对应的jvm参数为:

-Xms10g -Xmx10g -Xss512k -XX:PermSize=384m -XX:MaxPermSize=384m -XX:NewSize=7g -XX:MaxNewSize=7g
-XX:SurvivorRatio=8 -XX:MaxDirectMemorySize=4g -XX:+UseParNewGC -XX:ParallelGCThreads=4
-XX:MaxTenuringThreshold=9 -XX:+UseConcMarkSweepGC -XX:+DisableExplicitGC -XX:+UseCMSInitiatingOccupancyOnly
-XX:+ScavengeBeforeFullGC -XX:+UseCMSCompactAtFullCollection -XX:+CMSParallelRemarkEnabled
-XX:CMSFullGCsBeforeCompaction=9 -XX:CMSInitiatingOccupancyFraction=70

可以看到新生代为7G(其中Survivor为7G*1/10=700M),老年代为3G,对象最大的年龄居然只有9(-XX:MaxTenuringThreshold=9,导致new 区域进入到old区域的速度太快,old区域很快被撑满,频繁old gc),考虑到我这边的对象在一段时间内(几分钟)肯定会消亡,于是先进行一次调优尝试。

第一次调优

将年龄调到非常大,并且调大整个堆和young区,对应的jvm参数为

-Xms14g -Xmx14g -Xss512k -XX:PermSize=384m -XX:MaxPermSize=384m -XX:NewSize=12g -XX:MaxNewSize=12g 
-XX:SurvivorRatio=18 -XX:MaxDirectMemorySize=2g -XX:+UseParNewGC -XX:ParallelGCThreads=4 
-XX:MaxTenuringThreshold=30 -XX:+UseConcMarkSweepGC -XX:+DisableExplicitGC -XX:+UseCMSInitiatingOccupancyOnly 
-XX:+ScavengeBeforeFullGC -XX:+UseCMSCompactAtFullCollection -XX:+CMSParallelRemarkEnabled 
-XX:CMSFullGCsBeforeCompaction=9 -XX:CMSInitiatingOccupancyFraction=70 -XX:+CMSClassUnloadingEnabled 
-XX:SoftRefLRUPolicyMSPerMB=0 -XX:-ReduceInitialCardMarks -XX:+CMSPermGenSweepingEnabled 
-XX:CMSInitiatingPermOccupancyFraction=70

结果old区域一直为空,但是yong gc时间拉长许多,平均每次都要0.2s的时间,比之前的young gc多了整整三倍,这是无法接受的:

关于参数-XX:MaxTenuringThreshold:

在jdk1.7某个版本之前设置大于15表示是无穷大,即对象长命百岁,之后不管设置成多少,都是15,jdk1.8之后大于15直接报错。

第二次调优

将XX:MaxTenuringThreshold调整成15,由于之前cms 在old gc花的时间比较多,所以这里尝试serial old gc,对应的jvm参数为:

-Xms14g -Xmx14g -Xss512k -XX:PermSize=384m -XX:MaxPermSize=384m -XX:NewSize=12g -XX:MaxNewSize=12g 
-XX:SurvivorRatio=18 -XX:MaxDirectMemorySize=2g -XX:+UseParNewGC -XX:ParallelGCThreads=4 
-XX:MaxTenuringThreshold=15

发现效果确实比较明显,new gc的时间缩小了两倍左右,并且old gc的时间也从原来的15000ms 缩小到1500ms

第三次调优

仔细研究了一下第二种调优方式的组合,yong区域使用parNew 的方式,old区域使用serial old 的方式,如果在其他条件都相同的条件下使用parNew+cms的方式,old gc的时间会不会大幅度缩短?毕竟cms还是比较先进的收集器,于是分析了一下cms的几个阶段,有两个阶段是stop the world的,一个是初始化标记(intial mark),一个是重复标记(remark),重复标记的原因是并发标记阶段可能很多对象的状态都产生了变化,所以这个时候需要再次暂停用户所有的线程。

查看gc日志,发现remark的时间比较长,日志上下文如下

2016-08-27T17:26:00.108+0800: 261980.463: [GC [1 CMS-initial-mark: 2202742K(3145728K)] 2235571K(9751808K), 0.0561400 secs] [Times: user=0.05 sys=0.00, real=0.05 secs]
2016-08-27T17:26:00.165+0800: 261980.520: [CMS-concurrent-mark-start]
2016-08-27T17:26:00.918+0800: 261981.273: [CMS-concurrent-mark: 0.753/0.754 secs] [Times: user=1.27 sys=0.12, real=0.76 secs]
2016-08-27T17:26:00.918+0800: 261981.273: [CMS-concurrent-preclean-start]
2016-08-27T17:26:00.949+0800: 261981.304: [CMS-concurrent-preclean: 0.028/0.031 secs] [Times: user=0.04 sys=0.01, real=0.03 secs]
2016-08-27T17:26:00.949+0800: 261981.304: [CMS-concurrent-abortable-preclean-start]
CMS: abort preclean due to time 2016-08-27T17:26:06.159+0800: 261986.514: [CMS-concurrent-abortable-preclean: 5.197/5.210 secs] [Times: user=8.31 sys=0.66, real=5.21 secs]
2016-08-27T17:26:06.160+0800: 261986.515: Application time: 5.9951640 seconds
2016-08-27T17:26:06.161+0800: 261986.516: [GC[YG occupancy: 4756927 K (6606080 K)]
2016-08-27T17:26:06.161+0800: 261986.516: [Rescan (parallel) , 18.1621480 secs]2016-08-27T17:26:24.323+0800: 262004.678: [weak refs processing, 0.0000750 secs
2016-08-27T17:26:24.323+0800: 262004.678: [class unloading, 0.0069760 secs
2016-08-27T17:26:24.330+0800: 262004.685: [scrub symbol table, 0.0040020 secs
2016-08-27T17:26:24.334+0800: 262004.689: [scrub string table, 0.0009240 secs] [1 CMS-remark: 2202742K(3145728K)] 6959670K(9751808K), 18.1769610 secs] [Times: user=71.37 sys=0.06, real=18.18 secs]
2016-08-27T17:26:24.338+08002016-08-27T17:26:24.338+0800: : 262004.693: Application time: 0.0002080 seconds
262004.693: [CMS-concurrent-sweep-start]
2016-08-27T17:26:24.347+0800: 262004.702: Application time: 0.0079820 seconds
2016-08-27T17:26:24.868+0800: 262005.223: Application time: 0.5186580 seconds
{Heap before GC invocations=23680 (full 19):
par new generation   total 6606080K, used 5899299K [0x0000000568000000, 0x0000000728000000, 0x0000000728000000)
eden space 5872128K,  99% used [0x0000000568000000, 0x00000006ce5d44b8, 0x00000006ce680000)
from space 733952K,   3% used [0x00000006ce680000, 0x00000006d01b4910, 0x00000006fb340000)
to   space 733952K,   0% used [0x00000006fb340000, 0x00000006fb340000, 0x0000000728000000)

重新标记居然用了18s(对应这一段GC日志:[ 1 CMS-remark: 2202742K(3145728K)] 6959670K(9751808K), 18.1769610 secs ],另外,由于是多核心CPU的服务器,所以Times的real要远小于user的值,这个remark阶段是多线程并行的)!重新标记的时候,old区域的大小是固定的(这里设置成old区域的70%),按理说每次remark的时间应该都差不多才对,但是查了很多cms old gc日志,发现高峰期和低峰期remark的时间相差太大,两者的区别也只有eden区域,因为我这里eden的设置是比较大的,高峰期的时候,一次cms old从初始化标记开始,到remark之间,即并发标记阶段,用户程序会和gc线程是并发执行的。另外,由于我们程序的TPS比较高,我们再看一下GC日志,整个并发阶段的时间跨度为从2016-08-27T17:26:00.165+0800到2016-08-27T17:26:24.330+0800,总计24秒左右,那么,这个过程中肯定会新产生大量的新的对象。

如果remark之前能把eden区域的对象都清理一遍,那remark时,需要扫描的对象将会少很多很多对吧?cms正好有这么个参数-XX:+CMSScavengeBeforeRemark,这玩意的意思是在remark之前,来一次yong gc,恰好满足我们的要求,加了这个参数之后,对应的jvm参数为:

-Xms14g -Xmx14g -Xss512k -XX:PermSize=384m -XX:MaxPermSize=384m -XX:NewSize=12g -XX:MaxNewSize=12g 
 -XX:SurvivorRatio=18 -XX:MaxDirectMemorySize=2g -XX:+UseParNewGC -XX:ParallelGCThreads=4 
 -XX:MaxTenuringThreshold=15  -XX:+CMSScavengeBeforeRemark -XX:+UseConcMarkSweepGC -XX:+DisableExplicitGC 
 -XX:+UseCMSInitiatingOccupancyOnly -XX:+ScavengeBeforeFullGC -XX:+UseCMSCompactAtFullCollection 
 -XX:+CMSParallelRemarkEnabled -XX:CMSFullGCsBeforeCompaction=9 -XX:CMSInitiatingOccupancyFraction=70 

效果如下:

发现old gc的时间缩小到原来cms的1/100,窃喜。接下来分析gc日志,来验证我的猜想

2016-08-28T10:46:19.667+0800: 68444.509: [GC [1 CMS-initial-mark: 1469519K(2097152K)] 1545525K(14050944K), 0.0869600 secs] [Times: user=0.09 sys=0.00, real=0.08 secs]
2016-08-28T10:46:19.755+0800: 68444.597: [CMS-concurrent-mark-start]
2016-08-28T10:46:20.418+0800: 68445.260: [CMS-concurrent-mark: 0.663/0.663 secs] [Times: user=1.47 sys=0.24, real=0.66 secs]
2016-08-28T10:46:20.418+0800: 68445.260: [CMS-concurrent-preclean-start]
2016-08-28T10:46:20.438+0800: 68445.280: [CMS-concurrent-preclean: 0.018/0.020 secs] [Times: user=0.04 sys=0.01, real=0.02 secs]
2016-08-28T10:46:20.438+0800: 68445.280: [CMS-concurrent-abortable-preclean-start]
2016-08-28T10:46:24.542+0800: 68449.384: [CMS-concurrent-abortable-preclean: 4.090/4.104 secs] [Times: user=8.60 sys=1.40, real=4.10 secs]
2016-08-28T10:46:24.543+0800: 68449.385: Application time: 4.7880220 seconds

// 这里在remark之前进行一次young gc
============yong gc开始===============
2016-08-28T10:46:24.544+0800: 68449.386: [GC[YG occupancy: 5803756 K (11953792 K)]{Heap before GC invocations=4646 (full 7):
... ...2016-08-28T10:46:24.544+0800: 68449.386: [GC2016-08-28T10:46:24.544+0800: 68449.386: [ParNew: 5803756K->70533K (11953792K), 0.0668770 secs] 7273276K->1542365K(14050944K), 0.0669610 secs] [Times: user=0.25 sys=0.01, real=0.07 secs]
... ...
========yong gc结束,开始remark==============

2016-08-28T10:46:24.611+0800: 68449.453: [Rescan (parallel) , 0.0392690 secs]
2016-08-28T10:46:24.650+0800: 68449.492: [weak refs processing, 0.0001190 secs
2016-08-28T10:46:24.650+0800: 68449.492: [class unloading, 0.0072200 secs]2016-08-28T10:46:24.658+0800: 68449.500: [scrub symbol table, 0.0083430 secs]2016-08-28T10:46:24.666+0800: 68449.508: [scrub string table, 0.0011760 secs] [1 CMS-remark: 1471831K(2097152K)] 1542365K(14050944K), 0.1264420 secs] [Times: user=0.42 sys=0.01, real=0.13 secs]
2016-08-28T10:46:24.671+0800: 68449.513: [CMS-concurrent-sweep-start]
2016-08-28T10:46:24.672+0800: 68449.514: Application time: 0.0018070 seconds
2016-08-28T10:46:26.388+0800: 68451.230: [CMS-concurrent-sweep: 1.714/1.717 secs] [Times: user=3.70 sys=0.58, real=1.72 secs]
2016-08-28T10:46:26.388+0800: 68451.230: [CMS-concurrent-reset-start]
2016-08-28T10:46:26.396+0800: 68451.238: [CMS-concurrent-reset: 0.007/0.007 secs] [Times: user=0.02 sys=0.00, real=0.01 secs]
2016-08-28T10:46:34.434+0800: 68459.276: Application time: 9.7600810 seconds

可以看到,在remark之前,来一次yong gc,eden区域从5803756k降到70533k,给这个清理效果点个赞。这些空间如果不消除,那么cms将会在这些空间上进行非常耗时的标记,最后再看看remark的时间([1 CMS-remark: 1471831K(2097152K)] 1542365K(14050944K), 0.1264420 secs]),降到0.1264420s,和原来相比,整整一百倍的提高。

总结:

最后,对于长连接,push一类的海量服务端应用,16G内存8核心,推荐的JVM参数如下(仅供参考):

jdk1.7

-Xms13g -Xmx13g -Xss512k -XX:PermSize=384m -XX:MaxPermSize=384m -XX:NewSize=12g -XX:MaxNewSize=12g 
-XX:SurvivorRatio=18 -XX:MaxDirectMemorySize=2g -XX:+UseParNewGC -XX:ParallelGCThreads=4 -XX:MaxTenuringThreshold=15 -XX:+CMSParallelRemarkEnabled -XX:+CMSScavengeBeforeRemark -XX:+UseConcMarkSweepGC
-XX:+DisableExplicitGC -XX:+UseCMSInitiatingOccupancyOnly -XX:CMSInitiatingOccupancyFraction=70 -XX:+ScavengeBeforeFullGC -XX:+UseCMSCompactAtFullCollection -XX:CMSFullGCsBeforeCompaction=9  -XX:+CMSClassUnloadingEnabled  -XX:CMSInitiatingPermOccupancyFraction=70 -XX:+ExplicitGCInvokesConcurrent  
-XX:+PrintGCDetails -XX:+PrintGCDateStamps -XX:+PrintGCApplicationConcurrentTime -XX:+PrintHeapAtGC 
-Xloggc:/data/applogs/heap_trace.txt -XX:-HeapDumpOnOutOfMemoryError 
-XX:HeapDumpPath=/data/applogs/HeapDumpOnOutOfMemoryError

jdk1.8

-Xms13g -Xmx13g -Xss512k -XX:MetaspaceSize=384m -XX:MaxMetaspaceSize=384m -XX:NewSize=11g -XX:MaxNewSize=11g -XX:SurvivorRatio=18 -XX:MaxDirectMemorySize=2g -XX:+UseParNewGC -XX:ParallelGCThreads=4 -XX:MaxTenuringThreshold=15 -XX:+UseConcMarkSweepGC -XX:+DisableExplicitGC -XX:+UseCMSInitiatingOccupancyOnly -XX:+ScavengeBeforeFullGC -XX:+CMSScavengeBeforeRemark -XX:+CMSParallelRemarkEnabled -XX:CMSInitiatingOccupancyFraction=70 -XX:+CMSClassUnloadingEnabled -XX:SoftRefLRUPolicyMSPerMB=0 -XX:-ReduceInitialCardMarks -XX:+CMSClassUnloadingEnabled -XX:+ExplicitGCInvokesConcurrent -XX:+PrintGCDetails -XX:+PrintGCDateStamps -XX:+PrintGCApplicationConcurrentTime -XX:+PrintHeapAtGC -Xloggc:/data/applogs/heap_trace.txt -XX:-HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=/data/applogs/HeapDumpOnOutOfMemoryError"
我来评几句
登录后评论

已发表评论数()

相关站点

热门文章