Re: partly solved IO performance problems in 2.5.40 when writing to DVD-RAM/ZIP

[Andrea Arcangeli <andrea@suse.de>]

On Sun, Oct 06, 2002 at 06:08:00PM +0200, Moritz Franosch wrote:
> Theoretically, the two devices (e.g. the 120 GB HDD and the DVD-RAM in
> nr. 2) are independent (at least when they are on a different IDE
> controller), so the read should perform as good as without the
> additional writing activity (31 seconds, as given in the column
> "expected"). But in reality the read takes 1.8 times as long. With
> 2.4.50, there could be performance gains up to a factor of 2.4 (test
> nr. 4). But that's _much_ better than 2.4.19-pre5 where performance
> gains of up to a factor of 14 (test nr. 2) were possible.

one of the reasons 2.5 works better is that it has more than one bdflush
thread (aka pdflush), so it can submit in parallel to all your storages
and you don't risk to hang on writes on the very slow device, despite
your main harddisk could be completely idle. This isn't easily fixable
in 2.4, it's one of the things 2.5 is there for ;)

Andrea