On 10/05/2015 09:36 AM, Sandro Tosi wrote:
On Mon, Oct 5, 2015 at 5:25 PM, Nikolaus Rath <[email protected]> wrote:
Well, it only happens when using matplotlib, so there's got to be some
connection.
this is one good connection indeed. does it happen only on py3k or on
py2 as well?
It happens on py2 as well. Here is an impressive example:
Current jessie system:
$ time /usr/bin/python test.py
real 0m2.385s
user 0m2.312s
sys 0m0.064s
Freshly bootstrapped wheezy chroot:
$ time /usr/bin/python ./test.py
real 0m0.351s
user 0m0.320s
sys 0m0.024s
Do you have a suggestion what I could try? At the moment I'm seeing even
trivial matplotlib scripts take much longer than they used to, while Python
itself starts as quick as ever. That's why I think it's connected to MPL.
you can strace with timing and following the child processes, with -f
-ttt (f.e. you might found that that lseek()ing a cached file is not
that an issue) at least you can pin point where the time is spent.
Unfortunately there is no single call that sticks out - except that the
Jessie version makes more than 10 times as many syscalls:
$ ls -lh wheezy.log jessie.log
-rw-r--r-- 1 nikratio nikratio 14M Oct 7 09:41 jessie.log
-rw-r--r-- 1 nikratio nikratio 1.2M Oct 7 09:44 wheezy.log
At the moment I still think that the delay is really due to the
accumulated time taken by all these syscalls.
I have put the complete testcase at
https://trialphaenergy.box.com/shared/static/atpj07qtvdlgwft4ppp0l6q0jmjfo2a7.xz
(sorry for the ugly link).
Are you able to reproduce this?
Best,
-Nikolaus
