Dne 13.3.2018 v 12:17 Zdenek Kabelac napsal(a):
Dne 13.3.2018 v 05:03 Al Viro napsal(a):
On Sun, Mar 11, 2018 at 02:02:03PM +0100, Zdenek Kabelac wrote:

I've noticed quite some drop of performance (around 15% in some cases) where
execution of lvm2 tests took longer time - and while tests itself should not
really load CPU system  - the overall running time just got bigger.

Running bisect game pointed clearly to this commit:


commit 44c02a2c3dc55835e9f0d8ef73966406cd805001
Author: Al Viro <v...@zeniv.linux.org.uk>
Date:   Thu Oct 5 12:59:44 2017 -0400

     dev_ioctl(): move copyin/copyout to callers


clear revert of this commit on top of 3266b5bd97eaa72793df0b6e5a106c69ccc166c4
(recent  ~4.16-rc4)  restored timing of tests back.

I'm not sure why - so at this moment this is just a patch causing 15% longer
running times on lvm2 test suite.

I find it very hard to believe, TBH.  It doesn't go anywhere near drivers/md;
could you profile the damn thing and see where does it spend that extra time?

It really doesn't touch anything on non-sockets...


Well I'm 100% sure it's this patch causing slowdown of the test runtime
(just reverting this case makes it faster).

But even though I've already tried to create some much more simplified reproducer I'm still unclear what's behind it.

lvm2 alone is using sockets via i.e. udev communication - so the possible reason could be that used slows down  'new node' events processing in some way.

So for now the most visible reproducer is to run test in lvm2 build tree
in directory tests/:   'make check_local T=lvconvert-mirror-basic-0.sh'

I do have 'perf record' grabs if the would help anything - but looking at them myself there are not some directly visible things.

As said -  User+Sys timing of the running test reported by 'time' command remains the same - just 'real' time gets longer - which means it's waiting for something much longer then before  - lvm2 typically waits till udev finishes it's scanning.


Here is my update - after some time of investigation - it seems the core trouble is actually hidden inside test monitoring tool which is effectively grabbing test output through a PF_UNIX socket stream - there are no other sockets. (and via add printk() used commands are 21519 & 21505 for dev_ioctl() call).

Also there is no influence caused by udev as since timing is same with disable udev.

So ATM I'm checking, why there is such a big change in this case in the test's is run via internal application which is collecting logs compare with direct execution.

One more thing that has surprised me is the influence of CPU governor - the difference between ondemand/performance governor is another 10 seconds.
(on a test taking slight more then 1 minute).



Reply via email to