Linux 下 进程运行时内部函数耗时的统计 工具：pstack,strace,perf trace,systemtap

简单记录一些 在linux下 统计进程内部函数运行耗时的统计工具，主要是用作性能瓶颈分析。当然以下工具除了pstack功能单一之外，其他的工具都非常强大，这里仅仅整理特定场景的特定用法，用作协同分析。

以下工具需要追踪具体的进程，如果想要打印信息更全，建议编译的时候将符号信息都编译到二进制文件之中,-g选项

• strace
  strace -tttT -f -p $pid -o $save_file_name 追踪指定进程内部所有线程调用到的系统调用运行耗时，单位是秒，将统计结果保存到save_file_name的文件之中

  复制代码[pid 35467] 1596874136.770976 <... clock_gettime resumed> {1479879, 802163984}) = 0 <0.000293> [pid 35466] 1596874136.770989 <... futex resumed> ) = 0 <0.000290> [pid 35462] 1596874136.770998 <... futex resumed> ) = 1 <0.000289> [pid 35460] 1596874136.771007 <... sched_yield resumed> ) = 0 <0.000286> [pid 35457] 1596874136.771016 <... clock_gettime resumed> {1479879, 802216718}) = 0 <0.000280> [pid 35454] 1596874136.771025 <... sched_yield resumed> ) = 0 <0.000275>

  1
  2
  3
  4
  5
  6
  7
  [pid 35467] 1596874136.770976 <... clock_gettime resumed> {1479879, 802163984}) = 0 <0.000293>
  [pid 35466] 1596874136.770989 <... futex resumed> ) = 0 <0.000290>
  [pid 35462] 1596874136.770998 <... futex resumed> ) = 1 <0.000289>
  [pid 35460] 1596874136.771007 <... sched_yield resumed> ) = 0 <0.000286>
  [pid 35457] 1596874136.771016 <... clock_gettime resumed> {1479879, 802216718}) = 0 <0.000280>
  [pid 35454] 1596874136.771025 <... sched_yield resumed> ) = 0 <0.000275>
  

• pstack
  pstack $pid 追踪正在运行的进程的调用栈

  复制代码Thread 445 (Thread 0x7f17e35fe700 (LWP 33120)): #0 0x00007f18638be945 in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0 #1 0x00007f18640a0cbc in __gthread_cond_wait (__mutex=<optimized out>, __cond=__cond@entry=0x13af91c8) at gthr-default.h:864 #2 std::condition_variable::wait (this=this@entry=0x13af91c8, __lock=...) at condition_variable.cc:53 #3 0x000000000065c21f in rocksdb::ThreadPoolImpl::Impl::BGThread (this=this@entry=0x13af9130, thread_id=thread_id@entry=29) at util/threadpool_imp.cc:196 #4 0x000000000065c5c5 in rocksdb::ThreadPoolImpl::Impl::BGThreadWrapper (arg=0x13afbaf0) at util/threadpool_imp.cc:306 #5 0x00007f18640a6f20 in execute_native_thread_routine_compat () at thread.cc:94 #6 0x00007f18638bae25 in start_thread () from /lib64/libpthread.so.0 #7 0x00007f18635e834d in clone () from /lib64/libc.so.6

  1
  2
  3
  4
  5
  6
  7
  8
  9
  10
  Thread 445 (Thread 0x7f17e35fe700 (LWP 33120)):
  #0  0x00007f18638be945 in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
  #1  0x00007f18640a0cbc in __gthread_cond_wait (__mutex=<optimized out>, __cond=__cond@entry=0x13af91c8) at gthr-default.h:864
  #2  std::condition_variable::wait (this=this@entry=0x13af91c8, __lock=...) at condition_variable.cc:53
  #3  0x000000000065c21f in rocksdb::ThreadPoolImpl::Impl::BGThread (this=this@entry=0x13af9130, thread_id=thread_id@entry=29) at util/threadpool_imp.cc:196
  #4  0x000000000065c5c5 in rocksdb::ThreadPoolImpl::Impl::BGThreadWrapper (arg=0x13afbaf0) at util/threadpool_imp.cc:306
  #5  0x00007f18640a6f20 in execute_native_thread_routine_compat () at thread.cc:94
  #6  0x00007f18638bae25 in start_thread () from /lib64/libpthread.so.0
  #7  0x00007f18635e834d in clone () from /lib64/libc.so.6
  

• perf trace
  sudo perf trace -p $pid --duration 50 --call-graph dwarf -o $save_file_name
  统计运行耗时超过50ms的系统调用，并打印该系统调用的calltrace，并将打印的结果保存到save_file_name的文件之中，一般用于追踪IO性能问题

  复制代码563.060 (61.970 ms): rocksdb:high0/33123 fdatasync(fd: 8396 ) = 0 [0xffff80e79cbff9dd] (/usr/lib64/libc-2.17.so) rocksdb::PosixWritableFile::Sync (test) rocksdb::WritableFileWriter::SyncInternal (test) rocksdb::WritableFileWriter::Sync (test) rocksdb::BuildTable (test) rocksdb::FlushJob::WriteLevel0Table (test) rocksdb::FlushJob::Run (test)

  1
  2
  3
  4
  5
  6
  7
  8
  9
  563.060 (61.970 ms): rocksdb:high0/33123 fdatasync(fd: 8396          ) = 0
  			  [0xffff80e79cbff9dd] (/usr/lib64/libc-2.17.so)
  			   rocksdb::PosixWritableFile::Sync (test)
  			   rocksdb::WritableFileWriter::SyncInternal (test)
  			   rocksdb::WritableFileWriter::Sync (test)
  			   rocksdb::BuildTable (test)
  			   rocksdb::FlushJob::WriteLevel0Table (test)
  			   rocksdb::FlushJob::Run (test)
  

• systemtap
  这本身是一个非常强大的黑科技工具，除了内核态的调试之外，拥有符号表的用户进程也能够进行调试
  如下stap脚本trace_function_time，抓取运行时进程内部指定函数的耗时情况，并将耗时结果打印出来

  复制代码#!/bin/stap global sends #打印出来的单位是微妙 probe process("test").function("rocksdb::DBImpl::GetImpl").return { sends <<< gettimeofday_us() - @entry(gettimeofday_us()) } probe timer.s(1) { #每隔一秒打印一次 print(ctime(gettimeofday_s())) print("n") print(@hist_log(sends)) delete sends }

  1
  2
  3
  4
  5
  6
  7
  8
  9
  10
  11
  12
  13
  14
  15
  #!/bin/stap
  global sends 
  
  #打印出来的单位是微妙
  probe process("test").function("rocksdb::DBImpl::GetImpl").return {
    sends <<< gettimeofday_us() - @entry(gettimeofday_us()) 
  }
  
  probe timer.s(1) { #每隔一秒打印一次
   print(ctime(gettimeofday_s()))
   print("n")
   print(@hist_log(sends))
   delete sends    
  }
  

  需要sudo权限，直接sudo ./trace_function_time即可

  复制代码Fri Aug 7 03:24:18 2020 value |-------------------------------------------------- count 0 | 0 1 | 0 2 | 182 4 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 21280 8 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 14666 16 |@@@@@@@@@@@@@@@ 6518 32 |@@@@@@@@@@ 4608 64 |@@@@@@@@ 3421 128 |@@@@@ 2148 256 |@@ 923 512 | 361 1024 | 194 2048 | 75 4096 | 30 8192 | 0 16384 | 0

  1
  2
  3
  4
  5
  6
  7
  8
  9
  10
  11
  12
  13
  14
  15
  16
  17
  18
  19
  Fri Aug  7 03:24:18 2020 
  value |-------------------------------------------------- count
      0 |                                                       0
      1 |                                                       0
      2 |                                                     182
      4 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@  21280
      8 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@                 14666
     16 |@@@@@@@@@@@@@@@                                     6518
     32 |@@@@@@@@@@                                          4608
     64 |@@@@@@@@                                            3421
    128 |@@@@@                                               2148
    256 |@@                                                   923
    512 |                                                     361
   1024 |                                                     194
   2048 |                                                      75
   4096 |                                                      30
   8192 |                                                       0
  16384 |                                                       0
  

  打印的一秒内，统计处于各个微妙时间段内的请求个数

最后

以上就是贪玩海燕最近收集整理的关于Linux 下 进程运行时内部函数耗时的统计 工具：pstack,strace,perf trace,systemtap的全部内容，更多相关Linux内容请搜索靠谱客的其他文章。