kernel function probe overhead

[Aubrey Li]

Hi list,

I'm trying to profiling some Linux kernel subsystems by systemtap. So
I want to know the overhead of various kinds of systemtap probes.

Basically I sampled two kernel function probes,

probe kernel.function(arch_cpu_idle_enter).call
probe kernel.function(local_touch_nmi).call

========
a snip objdump of the code:

ffffffff81035740 <arch_cpu_idle_enter>:
ffffffff81035740:       e8 4b 76 55 00          callq
ffffffff8158cd90 <__fentry__>
ffffffff81035745:       e8 f6 96 ff ff          callq
ffffffff8102ee40 <local_touch_nmi>
========

because local irq is disabled, I suppose the duration of these two
probes to be ZERO.

However, I run the attached stp script and here is what I got:

$ stap test.stp
Press ^C to stop
Total time: 10106 miliseconds
t_end @count=2908 @min=676 @max=88152 @sum=15005802 @avg=5160

the duration varies from 676 to 88152, I understand all the tracing
mechanism should have overhead, while 676ns is acceptable, but 88152
is not.

And I have no idea why its range is so big?

Please let me know if I can provide more information on my side.

Thanks,
-Aubrey

$ stap -V
Systemtap translator/driver (version 3.1/0.167, commit
release-3.0-273-g7d27ff5084d6)
Copyright (C) 2005-2016 Red Hat, Inc. and others
This is free software; see the source for copying conditions.
tested kernel versions: 2.6.18 ... 4.8
enabled features: NLS

$ uname -a
Linux aubrey-hp 4.8.10+ #5 SMP Fri Nov 25 02:11:32 CST 2016 x86_64 GNU/Linux
-------------- next part --------------
A non-text attachment was scrubbed...
Name: test.stp
Type: application/octet-stream
Size: 527 bytes
Desc: not available
URL: <http://sourceware.org/pipermail/systemtap/attachments/20161125/ed92632c/attachment.obj>