背景
resin 是一款java实现的web服务器，在使用resin提供服务的时候发现：在低流量的情况下偶尔会出现高负载的情况
挑战的点：没有流量但是有负载抖动
先说结论
- 用户resin 网络线程配置过大(配置了2000)
- resin网络线程的行为是 poll -> accept
- 单个请求到达时，所有网络线程由于poll返回被唤醒，但是只有一个线程能够调用accept成功，类似于惊群
- 由于3，导致突增很多R进程，导致负载升高
确认是R进程导致的过程
查看D进程数量
部署/tmp/cpu.sh
#!/bin/bash
date 2>&1 >> /tmp/pid_D.log
/usr/bin/ps -aux | awk '$8 == "D"' >> /tmp/pid_D.log
在crontab中每分钟执行一次
查看/tmp/pid_D.log，几乎没有D进程，得出结论：当前loadavg主要由R构成
查看R进程调度情况
r进程是被唤醒的进程，通过ps查看属于一个采样动作，不准确，于是打开内核trace点，通过trace点能够100%采集唤醒动作
#!/bin/bash
echo 'comm == "java"' > /sys/kernel/debug/tracing/events/sched/sched_wakeup_new/filter
echo 'comm == "java"' > /sys/kernel/debug/tracing/events/sched/sched_wakeup/filter
echo 1 > /sys/kernel/debug/tracing/events/sched/sched_wakeup_new/enable
echo 1 > /sys/kernel/debug/tracing/events/sched/sched_wakeup/enable
nohup cat /sys/kernel/debug/tracing/trace_pipe > wakeup_trace.log &
打开sched_wakeup_new sched_wakeup两个trace点，输出到wakeup_trace.log中
使用work.py分析输出文件，按照固定时间段统计被唤醒的进程个数
from datetime import datetime
def calc_date(t):
boot_time = 1666330944.762007849
"""
date | cat > /dev/null # Make sure programs are in cache
set -- $(date +%s.%N) $(cat /proc/uptime)
echo "$1-$2" | bc
"""
return datetime.fromtimestamp(boot_time + t)
def main():
input = 'wakeup_trace.log'
index = 0
s = 0
interval = 50 #ms
dic = {}
diff = 0
for line in open(input):
try:
arr = line.split(':')
#print line.split()
pid = line.split()[6]
t = int(float(arr[0].split(' ')[-1])*1000)
cindex = t / interval
t1 = cindex*interval/1000.
if cindex != index:
print calc_date(t1), ' ', t1, 'per', interval, 'ms:', s, ' diff cnt: ', diff
s = 0
diff = 0
index = cindex
dic = {}
else:
if not dic.get(pid, False):
dic[pid] = True
diff += 1
s += 1
except Exception, e:
# print e
pass
if '__main__' == __name__:
main()
以50ms为周期统计, 输出到sched_a_out文件中
通过这样的手段分析出确实是进程过多唤醒导致高负载
现象复现
- 下载resin
wget https://caucho.com/download/resin-pro-4.0.66.tar.gz
- 在resin-pro-4.0.66/webapps/ROOT/下实现一个测试用的页面time.jsp
<html><body><center>
Now time is: <%=new java.util.Date()%>
</center></body></html>
- 在resin-pro-4.0.66目录执行下面的命令启/停resin
java -jar lib/resin.jar start
java -jar lib/resin.jar stop
- 通过修改conf/resin.properties中的配置项来修改监听线程数量，如下下是官方配置
# Throttle the number of active threads for a port
port_thread_max : 256
accept_thread_max : 32
accept_thread_min : 4
# keepalive_max : 512
- 修改配置将accept_thread_max accept_thread_min都调整到比较大的数值，比如200
# Throttle the number of active threads for a port
port_thread_max : 256
accept_thread_max : 200
accept_thread_min : 200
# keepalive_max : 512
- 使用ab进行压测， -c 参数使用1，表示同时只有一个请求发生，用户就是在这种低流量的场景发现负载升高，我们看看能否复现
ab -n 10000000 -c 1 http://127.0.0.1:8080/time.jsp
- uptime查看负载确认复现（当前使用的测试机为2c2g)
[root@VM-0-13-centos ~]# uptime
11:01:41 up 30 days, 12:20, 2 users, load average: 57.34, 15.76, 5.40
8.修改配置将accept_thread_max accept_thread_min都调整到1 负载降低下来
[root@VM-0-13-centos ~]# uptime
11:06:27 up 30 days, 12:24, 2 users, load average: 5.04, 17.69, 10.36
原因分析
strace角度
- 在使用200线程时
- 找到“-server com.caucho.server.resin.Resin” 以这个为参数的java进程进行strace，关注accept和poll
strace -p 1016685 -f -yy -e accept,poll- 可以看到在200线程时，有很多线程被poll唤醒, 有很多accpet调用，但是只用一个accept成功接受到socket
[pid 1016758] <... poll resumed>) = 1 ([{fd=45, revents=POLLIN}]) [pid 1016757] <... poll resumed>) = 1 ([{fd=45, revents=POLLIN}]) [pid 1016756] <... poll resumed>) = 1 ([{fd=45, revents=POLLIN}]) [pid 1016755] <... poll resumed>) = 1 ([{fd=45, revents=POLLIN}]) [pid 1016754] <... poll resumed>) = 1 ([{fd=45, revents=POLLIN}]) ...有很多accept [pid 1016733] accept(45<TCPv6:[[::]:8080]>, <unfinished ...> [pid 1016732] accept(45<TCPv6:[[::]:8080]>, <unfinished ...> [pid 1016731] accept(45<TCPv6:[[::]:8080]>, <unfinished ...> [pid 1016730] accept(45<TCPv6:[[::]:8080]>, <unfinished ...> [pid 1016729] accept(45<TCPv6:[[::]:8080]>, <unfinished ...> [pid 1016724] accept(45<TCPv6:[[::]:8080]>, <unfinished ...> [pid 1016723] accept(45<TCPv6:[[::]:8080]>, <unfinished ...> [pid 1016721] accept(45<TCPv6:[[::]:8080]>, <unfinished ...> [pid 1016720] accept(45<TCPv6:[[::]:8080]>, <unfinished ...>只有一个accept调用成功 [pid 1017179] <... accept resumed>{sa_family=AF_INET6, sin6_port=htons(37432), sin6_flowinfo=htonl(0), inet_pton(AF_INET6, "::ffff:127.0.0.1", &sin6_addr), sin6_scope_id=0}, [28]) = 60<TCPv6:[[::ffff:127.0.0.1]:8080->[::ffff:127.0.0.1]:37432]>其他都失败 [pid 1016967] <... accept resumed>0x7fd965352100, [28]) = -1 EAGAIN (资源暂时不可用) [pid 1016966] <... accept resumed>0x7fd965453340, [28]) = -1 EAGAIN (资源暂时不可用) [pid 1016964] <... accept resumed>0x7fd965654f80, [28]) = -1 EAGAIN (资源暂时不可用) ... - 在使用单个监听线程的时候, 可以看到每一个成功的poll都对应一个成功的accept
[pid 1019707] poll([{fd=59<TCPv6:[[::ffff:127.0.0.1]:8080->[::ffff:127.0.0.1]:39736]>, events=POLLIN|POLLERR}], 1, 120000) = 1 ([{fd=59, revents=POLLIN}]) [pid 1019697] poll([{fd=45<TCPv6:[[::]:8080]>, events=POLLIN|POLLERR}], 1, -1) = 1 ([{fd=45, revents=POLLIN}]) [pid 1019697] accept(45<TCPv6:[[::]:8080]>, {sa_family=AF_INET6, sin6_port=htons(39738), sin6_flowinfo=htonl(0), inet_pton(AF_INET6, "::ffff:127.0.0.1", &sin6_addr), sin6_scope_id=0}, [28]) = 59<TCPv6:[[::ffff:127.0.0.1]:8080->[::ffff:127.0.0.1]:39738]>
从以上的信息来看，是因为resin的设计是“accept_thread_max accept_thread_min控制同时poll的线程数量，每一个poll成功之后会调用accept”，这就导致如果线程数量过多会造成“小流量唤醒很多线程的poll调用，但是真正只有很少的accept调用成功”，可以看成是一种“惊群”的现象。
用户配置2000线程的原因
用户误以为resin的网络线程的处理方式是串行的，也就是“在一个请求未处理完之前不会处理下一个请求“，这个认知已经通过如下的实验帮助用户纠正过来。
- 使用一个监听线程
- ab -c使用20进行压测
从输出可以看出, 单个监听线程每秒可以扛住4649个request，所以一定不是串行模式，而应该是多路io复用的模式[root@VM-0-13-centos ~]# ab -n 1000 -c 20 http://127.0.0.1:8080/time.jsp ... Concurrency Level: 20 Time taken for tests: 0.215 seconds Complete requests: 1000 Failed requests: 0 Total transferred: 348000 bytes HTML transferred: 91000 bytes Requests per second: 4649.93 [#/sec] (mean) Time per request: 4.301 [ms] (mean) Time per request: 0.215 [ms] (mean, across all concurrent requests) Transfer rate: 1580.25 [Kbytes/sec] received - 真正影响业务并发线程数的参数是
# port_thread_max restricts the active threads available for a single port. port_thread_max : 256
对于高负载定位的快捷办法
insmod loadavg_tracer.ko load_threshold=20 dump_interval=5
上面的命令会在load高于20的时候打印R/D进程的栈
[2640203.171984] R [001] resin-port-8080 1025456
[2640203.172912]
poll_schedule_timeout.constprop.17+0x46/0x70
do_sys_poll+0x46b/0x500
__x64_sys_poll+0x3b/0x150
do_syscall_64+0x4d/0x150
entry_SYSCALL_64_after_hwframe+0x44/0xa9
[2640203.177809] R [001] resin-port-8080 1025457
[2640203.178725]
poll_schedule_timeout.constprop.17+0x46/0x70
do_sys_poll+0x46b/0x500
__x64_sys_poll+0x3b/0x150
do_syscall_64+0x4d/0x150
entry_SYSCALL_64_after_hwframe+0x44/0xa9
[2640203.183616] R [000] resin-port-8080 1025459
[2640203.184535]
poll_schedule_timeout.constprop.17+0x46/0x70
do_sys_poll+0x46b/0x500
__x64_sys_poll+0x3b/0x150
do_syscall_64+0x4d/0x150
entry_SYSCALL_64_after_hwframe+0x44/0xa9
[2640203.189394] R [001] resin-port-8080 1025460
[2640203.190102]
poll_schedule_timeout.constprop.17+0x46/0x70
do_sys_poll+0x46b/0x500
__x64_sys_poll+0x3b/0x150
do_syscall_64+0x4d/0x150
entry_SYSCALL_64_after_hwframe+0x44/0xa9
[2640203.193815] R [001] resin-port-8080 1025461
[2640203.194624]
poll_schedule_timeout.constprop.17+0x46/0x70
do_sys_poll+0x46b/0x500
__x64_sys_poll+0x3b/0x150
do_syscall_64+0x4d/0x150
entry_SYSCALL_64_after_hwframe+0x44/0xa9
[2640203.200751] R [001] resin-port-8080 1025511
[2640203.201538]
timer_callback.cold.2+0x154/0x1c8 [loadavg_tracer]
__hrtimer_run_queues+0x117/0x2a0
hrtimer_interrupt+0x101/0x210
smp_apic_timer_interrupt+0x74/0x140
apic_timer_interrupt+0xf/0x20
finish_task_switch+0x8b/0x250
poll_schedule_timeout.constprop.17+0x46/0x70
do_sys_poll+0x46b/0x500
__x64_sys_poll+0x3b/0x150
do_syscall_64+0x4d/0x150
entry_SYSCALL_64_after_hwframe+0x44/0xa9
1720
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
