Re: bdflush synchronous IO on prerelease-diff

[Marcelo Tosatti]

On Wed, 3 Jan 2001, Marcelo Tosatti wrote:

> 
> Hi Linus, 
> 
> I've noticed you changed bdflush to do synchronous IO on page_launder().
> 
> That seems to be a performance problem, since kflushd will have to wait
> for dirty buffers to get synced instead looping on the inactive dirty
> list more often. 
> 
> Here is a patch to change this. 
> 
> --- linux.orig/fs/buffer.c      Wed Jan  3 22:43:24 2001
> +++ linux/fs/buffer.c   Thu Jan  4 00:28:50 2001
> @@ -2710,7 +2710,7 @@
>  
>                 flushed = flush_dirty_buffers(0);
>                 if (free_shortage())
> -                       flushed += page_launder(GFP_KERNEL, 1);
> +                       flushed += page_launder(GFP_KERNEL, 0);
>  
>                 /*
>                  * If there are still a lot of dirty buffers around,

And here are some dbench numbers:

12 clients / 64M ram
without the patch: 10.04 Mb/s  
with the patch: 14.47 Mb/s

48 clients / 256M ram 
without the patch: 14.40 Mb/s
with the patch: 15.69 Mb/s