Welcome to WuJiGu Developer Q&A Community for programmer and developer-Open, Learning and Share
Welcome To Ask or Share your Answers For Others

Categories

0 votes
761 views
in Technique[技术] by (71.8m points)

linux - High Kernel CPU when running multiple python programs

I developed a python program that does heavy numerical calculations. I run it on a linux machine with 32 Xeon CPUs, 64GB RAM, and Ubuntu 14.04 64-bit. I launch multiple python instances with different model parameters in parallel to use multiple processes without having to worry about the global interpreter lock (GIL). When I monitor the cpu utilization using htop, I see that all cores are used, however most of the time by kernel. Generally, the kernel time is more than twice the user time. I'm afraid that there is a lot of overhead going on on the system level, but I'm not able to find the cause for this.

How would one reduce the high kernel CPU usage?

Here are some observation I made:

  • This effect appears independent of whether I run 10 jobs or 50. If there are fewer jobs than cores, not all cores are used, but the ones that are used still have a high CPU usage by the kernel
  • I implemented the inner loop using numba, but the problem is not related to this, since removing the numba part does not resolve the problem
  • I also though that it might be related to using python2 similar to the problem mentioned in this SO question but switching from python2 to python3 did not change much
  • I measured the total number of context switches performed by the OS, which is about 10000 per second. I'm not sure whether this is a large number
  • I tried increasing the python time slices by setting sys.setcheckinterval(10000) (for python2) and sys.setswitchinterval(10) (for python3) but none of this helped
  • I tried influencing the task scheduler by running schedtool -B PID but this didn't help

Edit: Here is a screenshot of htop: enter image description here

I also ran perf record -a -g and this is the report by perf report -g graph:

Samples: 1M of event 'cycles', Event count (approx.): 1114297095227                                   
-  95.25%          python3  [kernel.kallsyms]                           [k] _raw_spin_lock_irqsave   ◆
   - _raw_spin_lock_irqsave                                                                          ?
      - 95.01% extract_buf                                                                           ?
           extract_entropy_user                                                                      ?
           urandom_read                                                                              ?
           vfs_read                                                                                  ?
           sys_read                                                                                  ?
           system_call_fastpath                                                                      ?
           __GI___libc_read                                                                          ?
-   2.06%          python3  [kernel.kallsyms]                           [k] sha_transform            ?
   - sha_transform                                                                                   ?
      - 2.06% extract_buf                                                                            ?
           extract_entropy_user                                                                      ?
           urandom_read                                                                              ?
           vfs_read                                                                                  ?
           sys_read                                                                                  ?
           system_call_fastpath                                                                      ?
           __GI___libc_read                                                                          ?
-   0.74%          python3  [kernel.kallsyms]                           [k] _mix_pool_bytes          ?
   - _mix_pool_bytes                                                                                 ?
      - 0.74% __mix_pool_bytes                                                                       ?
           extract_buf                                                                               ?
           extract_entropy_user                                                                      ?
           urandom_read                                                                              ?
           vfs_read                                                                                  ?
           sys_read                                                                                  ?
           system_call_fastpath                                                                      ?
           __GI___libc_read                                                                          ?
    0.44%          python3  [kernel.kallsyms]                           [k] extract_buf              ?
    0.15%          python3  python3.4                                   [.] 0x000000000004b055       ?
    0.10%          python3  [kernel.kallsyms]                           [k] memset                   ?
    0.09%          python3  [kernel.kallsyms]                           [k] copy_user_generic_string ?
    0.07%          python3  multiarray.cpython-34m-x86_64-linux-gnu.so  [.] 0x00000000000b4134       ?
    0.06%          python3  [kernel.kallsyms]                           [k] _raw_spin_unlock_irqresto?
    0.06%          python3  python3.4                                   [.] PyEval_EvalFrameEx       

It seems as if most of the time is spent calling _raw_spin_lock_irqsave. I have no idea what this means, though.

See Question&Answers more detail:os

与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…
Welcome To Ask or Share your Answers For Others

1 Answer

0 votes
by (71.8m points)

If the problem exists in kernel, you should narrow down a problem using a profiler such as OProfile or perf.

I.e. run perf record -a -g and than read profiling data saved into perf data using perf report. See also: linux perf: how to interpret and find hotspots.


In your case high CPU usage is caused by competition for /dev/urandom -- it allows only one thread to read from it, but multiple Python processes are doing so.

Python module random is using it only for initialization. I.e:

$ strace python -c 'import random;
while True:
    random.random()'
open("/dev/urandom", O_RDONLY)     = 4
read(4, "163636636}"..., 2500) = 2500
close(4)                                   <--- /dev/urandom is closed

You may also explicitly ask for /dev/urandom by using os.urandom or SystemRandom class. So check your code which is dealing with random numbers.


与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…
Welcome to WuJiGu Developer Q&A Community for programmer and developer-Open, Learning and Share
...