Re: Prereading block device doesn't help filesystems

[Chris Mason <mason@suse.com>]

On Tue, 2003-06-03 at 13:49, Mike Fedyk wrote:
> On Tue, Jun 03, 2003 at 09:33:36AM +0400, Oleg Drokin wrote:
> > (e.g. dd if=/dev/hdb1 /dev/zero bs=1024k count=500 (replace 500 with the half of your RAM
> > in megabytes) and you have large enough RAM, there is a chance that this will cache
> > enough data to keep stuff moving more quickly. 
> 
> That won't work either.
> 
> Basically you're copying the data into buffer cache, and the page cache
> won't be able to pick it up, so you're going to read from the disk twice.
> Even with Andrea's buffercache in pagecache from 2.4.10 it doesn't do the
> aliasing IIRC.  And there were some long threads where Linus said he
> wouldn't accept patches to do the aliasing either.
> 
> Don't know if he has changed his mind in 2.5 though...
> 

There are two caches used by reiserfs.  The page cache has file data,
and the buffer cache has FS metadata.  The buffer cache doesn't really
exist anymore, it's really just the page cache for the file
corresponding to the block device.

So, if you dd if=/dev/hdb1 of=/dev/zero, you are reading into the page
cache of the block device which means you are populating the metadata
cache for reiserfs as well.

With packed tails and many small files, this represents the bulk of the
data on the FS, and it should make a performance difference if you have
enough ram to make it worth while, and your access patterns keep the
blocks in cache.

-chris