Re: [PATCH] proc: speedup /proc/stat handling

[KAMEZAWA Hiroyuki]

On Mon, 23 Jan 2012 14:33:54 +0400
Glauber Costa <glommer@xxxxxxxxxxxxx> wrote:

> On 01/23/2012 02:16 PM, KAMEZAWA Hiroyuki wrote:
> > On Fri, 20 Jan 2012 16:59:24 +0100
> > Eric Dumazet<eric.dumazet@xxxxxxxxx>  wrote:
 >>
> >> An alternative to 1) would be to remember the largest m->count reached
> >> in show_stat()
> >>
> >
> > nice catch. But how about using usual seq_file rather than single_open() ?
> > I just don't like multi-page buffer for this small file...very much.
> >
> > A rough patch here, maybe optimization will not be enough. (IOW, this may be slow.)
> >
> 
> I myself don't like it very much, at least at first sight.
> Even with optimizations applied, I doubt we can make this approach 
> faster than what we currently do for /proc/stat.
> 
IIUC, most of cost comes from printing " 0".

> Also, the code gets a lot harder to read and grasp. Problem is, unlike 
> most of the stuff using seq_file, /proc/stat shows a lot of different 
> kinds of information, not a single kind of easily indexable information.
> 

I think current one is too simple. But yes, may not be worth to to use usual
seq_file sequence.

I did some optimization around number(). Because my environ is small.
size of /proc/stat is 2780.
[kamezawa@bluextal test]$ wc -c /proc/stat
2780 /proc/stat

Test program is this.
== test program (read 1000 tiems.)==

#!/usr/bin/python

num = 0

with open("/proc/stat") as f:
        while num < 1000 :
                data = f.read()
                f.seek(0, 0)
                num = num + 1


== Before patch (3.3-rc1) ==
[kamezawa@bluextal test]$ time ./stat_check.py

real    0m0.142s
user    0m0.022s
sys     0m0.117s

== After patch ==
[root@bluextal test]# time ./stat_check.py

real    0m0.096s
user    0m0.024s
sys     0m0.069s


==
In above, most of improvements comes from replacing seq_printf() with seq_puts().

If the number of cpu increases, the most costly one will be kstat_irqs().

perf record after patch:
    19.03%  stat_check.py  [kernel.kallsyms]    [k] memcpy
     7.83%  stat_check.py  [kernel.kallsyms]    [k] seq_puts
     6.83%  stat_check.py  [kernel.kallsyms]    [k] kstat_irqs
     5.75%  stat_check.py  [kernel.kallsyms]    [k] radix_tree_lookup
     5.24%  stat_check.py  libpython2.6.so.1.0  [.] 0x8ccb0
     4.35%  stat_check.py  [kernel.kallsyms]    [k] vsnprintf
     3.68%  stat_check.py  [kernel.kallsyms]    [k] sub_preempt_count
     3.45%  stat_check.py  [kernel.kallsyms]    [k] number

If we can find kstat_irqs()==0 without walking all possible cpus, we can
cut most of costs...

Anyway, this is my final version. I'll go to my usual work ;)

==