Re: [RFC] ext3 writepages for writeback mode

[Andrew Morton <akpm@osdl.org>]

Badari Pulavarty <pbadari@us.ibm.com> wrote:
>
> BEFORE: (without writepages support)
>                                                                                                        
> elm3b29:/mnt # touch file
> elm3b29:/mnt # time /tmp/writer file
> real    0m23.746s user    0m0.000s sys     0m5.020s (allocation)
> elm3b29:/mnt # time /tmp/writer file
> real    0m20.950s user    0m0.001s sys     0m2.278s (no alloc)
> elm3b29:/mnt # time /tmp/writer file
> real    0m21.030s user    0m0.001s sys     0m2.254s (no alloc)
> elm3b29:/mnt # time /tmp/writer file
> real    0m20.577s user    0m0.001s sys     0m2.184s (no alloc)
>                                                                                                        
> ====
> AFTER: (with writepages support)
>                                                                                                        
> elm3b29:/mnt # touch file
> elm3b29:/mnt # time /tmp/writer file
> real    0m23.230s user    0m0.001s sys     0m4.132s (allocation)
> elm3b29:/mnt # time /tmp/writer file
> real    0m20.175s user    0m0.004s sys     0m1.756s (no alloc)
> elm3b29:/mnt # time /tmp/writer file
> real    0m20.368s user    0m0.001s sys     0m1.696s (no alloc)
> elm3b29:/mnt # time /tmp/writer file
> real    0m20.626s user    0m0.002s sys     0m1.763s (no alloc)

Holy cow.  I'm shocked.

There's no system CPU time involved, and the user CPU time didn't change. 
We must be getting better I/O scheduling for some reason.  I wonder what it
is?

That, or we're forgetting to write something ;)

What journalling mode were you using?
What I/O scheduler?
What sort of disk system?