https://blogs.Oracle.com/Database4CN/entry/%E5%85%B3%E4%BA%8Esys_cpu_usage_100_%E9%97%AE%E9%A2%98%E7%9A%84%E5%88%86%E6%9E%90
SNAP_INTERVAL 15CPU_COUNT 32zzz ***Fri Dec 2 17:05:03 CST 2016procs -----------memory---------- ---swap-- -----io---- --system-- -----cpu------r b swpd free buff cache si so bi bo in cs us sy id wa st48 0 0 22026868 213392 37138888 0 0 21 31 13 39 6 8 86 0 044 1 0 21968452 213392 37138900 0 0 0 360 1093 537 8 92 0 0 044 1 0 21941632 213392 37139028 0 0 0 288 1080 371 9 91 0 0 0......zzz ***Fri Dec 2 17:10:12 CST 2016procs -----------memory---------- ---swap-- -----io---- --system-- -----cpu------r b swpd free buff cache si so bi bo in cs us sy id wa st121 2 0 21495944 218356 37142412 0 0 21 31 13 39 6 9 85 0 0122 4 0 21486192 218356 37142432 0 0 0 308 119 753 7 93 0 0 0121 2 0 21478868 218364 37142424 0 0 0 592 97 517 5 95 0 0 0首先我看了一遍客户提供的AWR,发现DB相当空闲,虽然CPU time占据了91.2,但是总的CPU Time在119分钟的采样中只有18345秒(305分钟),相对于客户32个CPU Core来说不是个问题。Snap Id Snap Time Sessions Cursors/SessionBegin Snap: 38119 02-Dec-16 16:00:28 255 63.2End Snap: 38121 02-Dec-16 18:00:18 193 48.7Elapsed: 119.83 (mins)DB Time: 335.30 (mins) <<< ------------相当空闲Top 5 Timed EventsEvent Waits Time(s) Avg Wait(ms) % Total Call Time Wait ClassCPU time 18,345 91.2 os thread startup 971 937 965 4.7 Concurrencylatch free 582 657 1,128 3.3 Otherdb file sequential read 4,712,799 345 0 1.7 User I/Olog file parallel write 247,562 258 1 1.3 System I/OAWR中没发现什么异常,DB的alert log显示一些无法fork进程的消息,估计是资源紧张了。Fri Dec 2 17:06:16 2016Process q002 died, see its trace fileFri Dec 2 17:06:16 2016ksvcreate: Process(q002) creation failed好吧,一般情况下如果我们发现CPU高,无论是sys的还是user的,我们一般的做法是先定位top function call然后通过这些function call来定位oracle或者OS行为,并且通过这些call来搜索与匹配已知问题。在linux上,最方便收集这些信息的就是用perf这个工具。关于perf,参见:https://perf.wiki.kernel.org/index.php/Tutorial结果客户说他们无法安装perf命令,不过他提到他的OS中显示很多错误:Dec 2 17:05:23 erpdb1 kernel: BUG: soft lockup - CPU#5 stuck for 10s! [oracle:15668]Dec 2 17:05:23 erpdb1 kernel: CPU 5:Dec 2 17:05:23 erpdb1 kernel: Call Trace:Dec 2 17:05:23 erpdb1 kernel: [<ffffffff8000e9a8>] __set_page_dirty_nobuffers+0xc2/0xe9Dec 2 17:05:23 erpdb1 kernel: [<ffffffff80007c1b>] unmap_vmas+0x522/0x904Dec 2 17:05:23 erpdb1 kernel: [<ffffffff80012d08>] unmap_region+0xb8/0x12bDec 2 17:05:23 erpdb1 kernel: [<ffffffff80011e45>] do_munmap+0x21b/0x29aDec 2 17:05:23 erpdb1 kernel: [<ffffffff800655ab>] __down_write_nested+0x12/0x92Dec 2 17:05:23 erpdb1 kernel: [<ffffffff80121e88>] sys_shmdt+0x5b/0x133Dec 2 17:05:23 erpdb1 kernel: [<ffffffff8005e28d>] tracesys+0xd5/0xe0通过call stack,看来在回收内存时报错了,推测这个错误应当发生在进程退出阶段,不过难以断定这些错误与sys cpu高的因果关系。结合客户描述的现象,这看起来很像连接风暴,因此我们检查了ps的输出,发现进程数并未明显增加,不过问题最严重的时间断片了。这些零碎的信息并不能给我们一个很清晰的线索。$ awk '/$ORACLE_SID/{n++;next}/^zzz/{if(t)print t,"-",n;t=$0;n=0}END{print t,"-",n}' XXXX_ps_16.12.02.1700.datzzz ***Fri Dec 2 17:04:18 CST 2016 - 235zzz ***Fri Dec 2 17:04:33 CST 2016 - 236zzz ***Fri Dec 2 17:04:48 CST 2016 - 229zzz ***Fri Dec 2 17:05:03 CST 2016 - 228 <<<< 此时问题实际上已经发生了zzz ***Fri Dec 2 17:05:19 CST 2016 - 178 <<<< 17:05 ~ 17:13 的断片了zzz ***Fri Dec 2 17:13:19 CST 2016 - 283 <<<<zzz ***Fri Dec 2 17:13:34 CST 2016 - 283zzz ***Fri Dec 2 17:13:49 CST 2016 - 196接下来看了top,发现虽然OS的sys CPU高,不过top的process都是oracle,表明此问题一定与oracle有点关系。zzz ***Fri Dec 2 17:05:03 CST 2016top - 17:05:05 up 9:24, 3 users, load average: 41.76, 28.54, 19.68Tasks: 660 total, 45 running, 615 sleeping, 0 stopped, 0 zombieCpu(s): 8.3%us, 91.7%sy, 0.0%ni, 0.0%id, 0.0%wa, 0.0%hi, 0.0%si, 0.0%stMem: 65993408k total, 44046040k used, 21947368k free, 213392k buffersSwap: 62918564k total, 0k used, 62918564k free, 37139028k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND19610 oracle 25 0 9917m 8.5g 8.5g R 101.8 13.6 1:19.76 oracle19756 oracle 25 0 9917m 7.0g 7.0g R 100.9 11.1 1:05.76 oracle19760 oracle 25 0 9917m 6.7g 6.7g R 100.9 10.7 1:06.56 oracle19942 oracle 25 0 9917m 5.1g 5.1g R 100.9 8.0 0:46.67 oracle20107 oracle 25 0 9917m 3.1g 3.1g R 100.9 4.9 0:26.39 oracle20204 oracle 25 0 9917m 1.2g 1.2g R 100.9 1.9 0:10.63 oracle19486 oracle 25 0 9917m 9.3g 9.3g R 99.9 14.8 1:25.10 oracle19721 oracle 25 0 9917m 6.9g 6.8g R 99.9 10.9 1:08.22 oracle那么问题来了,oracle软件一般都是执行user code,因此大多数情况下会消耗user space的CPU,怎么会消耗sys CPU呢? 先man一下top:sy - This is the amount of time that the CPU spent running the kernel.All the processes and system resources are handled by the Linux kernel.When a user space process needs something from the system, for examplewhen it needs to allocate memory, perform some I/O, or it needs tocreate a child process, then the kernel is running.这说明oracle进程是有可能消耗kernel space的CPU的,比如申请内存,执行I/O等。挑出上面的top列出的进程,在ps输出中找规律:
$ grep 19610 Dec*TIME USER PID PPID PRI %CPU %MEM VSZ rss WCHAN S STARTED TIME COMMANDDec 2 17:03:17 oracle 19610 1 14 78.5 0.9 10155628 605564 - R 17:03:10 00:00:05 ora_q002_XXXXDec 2 17:03:32 oracle 19610 1 14 57.6 2.1 10155628 1437940 - R 17:03:09 00:00:13 ora_q002_XXXXDec 2 17:03:47 oracle 19610 1 14 67.6 4.3 10155628 2868692 - R 17:03:10 00:00:25 ora_q002_XXXXDec 2 17:04:02 oracle 19610 1 14 75.8 6.9 10155628 4559708 - R 17:03:09 00:00:40 ora_q002_XXXXDec 2 17:04:18 oracle 19610 1 14 76.7 9.1 10155628 6015688 - R 17:03:10 00:00:52 ora_q002_XXXXDec 2 17:04:33 oracle 19610 1 14 70.9 10.4 10155628 6865876 - R 17:03:09 00:00:59 ora_q002_XXXXDec 2 17:04:48 oracle 19610 1 14 67.7 11.6 10155628 7684088 - R 17:03:09 00:01:07 ora_q002_XXXXDec 2 17:05:03 oracle 19610 1 14 68.9 13.3 10155628 8838576 - R 17:03:10 00:01:18 ora_q002_XXXX$ grep 19756 Dec*TIME USER PID PPID PRI %CPU %MEM VSZ RSS WCHAN S STARTED TIME COMMANDDec 2 17:03:47 oracle 19756 1 16 50.0 0.3 10155628 222508 - R 17:03:44 00:00:02 oracleXXXX (LOCAL=NO)Dec 2 17:04:02 oracle 19756 1 14 47.9 1.4 10155628 961764 - R 17:03:43 00:00:09 oracleXXXX (LOCAL=NO)Dec 2 17:04:18 oracle 19756 1 14 55.5 3.0 10155628 2021664 - R 17:03:44 00:00:18 oracleXXXX (LOCAL=NO)Dec 2 17:04:33 oracle 19756 1 14 68.4 5.6 10155628 3703572 - R 17:03:43 00:00:34 oracleXXXX (LOCAL=NO)Dec 2 17:04:48 oracle 19756 1 14 75.4 8.2 10155628 5459948 - R 17:03:43 00:00:49 oracleXXXX (LOCAL=NO)Dec 2 17:05:03 oracle 19756 1 14 80.6 10.9 10155628 7217680 - R 17:03:44 00:01:04 oracleXXXX (LOCAL=NO)从以上输出可以发现一个明显规律: 这些进程的RSS在1分多钟从几十M变成7~8G,但是VSZ却没有变化。接着man psVSZ: virtual memory usage of entire process. vm_lib + vm_exe + vm_data + vm_stackRSS: resident set size, the non-swapped physical memory that a task has used (in kiloBytes). (alias rssize, rsz).任何一个oracle进程的VSZ约等于SGA加上这个进程的PGA(实际上VSZ还包含一些kernel内存),正常情况下一个进程的pga是很小的。以上输出中VSZ没有改变,因此发生巨大变化的RSS申请的内存一定不是PGA而是SGA(因为如果增长的是PAG那么VSZ也会跟着增长)。好吧,那么只有一个可能了,那就是这个进程在touch整个sga,为什么会这样?们需要再回到原点再看一眼AWR的数据库参数信息,赫然发现如下内容:sga_max_size 10250878976sga_target 8304721920pre_page_sga TRUE <<<<------------------看这里这个设置中的sga_max_size正好10g,与我们在ps中看到的VSZ正好相等。问题的原因是客户设置了pre_page_sga=true,这样在oracle进程启动阶段会touch整个SGA,这个过程中会调用OS的sys call来touch 整个 shared memory entry,因此引发了高SYS CPU消耗。参见如下文档的描述:
Health Check Alert: Consider setting PRE_PAGE_SGA to FALSE (Doc ID 957525.1) 回过头来再看alert log,观察参数pre_page_sga是什么时候改的,发现它在很久以前的很多次重启就是true了。也就是说,这个问题一直都存在,只是客户最近维护UPS之后才发现,维护UPS这个信息误导了我们。
新闻热点
疑难解答