On Thu, Mar 18, 2004 at 12:09:45PM -0700, Pat LaVarre wrote: > Matt D: > > >>>Subject: Re: [usb-storage] Re: [linux-usb-devel] unneeded subclass > >>>error in driv er > >>>Try changing max_sectors in linux/drivers/usb/storage/scsiglue.c to a > >>>smaller number (i.e. 16 or 32 or 64) and see if that improves the > >>>transfer speed. > >> > >>Cool, 32 or 64, about 5x faster than when i went to 128. > > > >Given results like this, I wonder if we shouldn't reduce max_sectors > >permanently... > > Urgh. > > You're just kidding, yes?
No, I'm not kidding.
My reasoning goes like this:
(1) smaller fragments improves interoperability. Interop is good.
(2) we periodically get reports of devices like the one above -- a big drop
in performance with large transfers. My guess is an internal buffer isn't
very big. Since 2.4 used smaller fragments by default than 2.6, they
see this as a performance drop with the new version.
We should let people who want more performance tweak the fragment size up
(to a reasonably large limit) if they want to try -- this could lead to
device failures or better performance. Letting people shoot themselves in
the foot is a long-standing tradition in Linux, but I don't want to leave
the 'non-power users' out in the cold.
Matt
--
Matthew Dharm Home: [EMAIL PROTECTED]
Maintainer, Linux USB Mass Storage Driver
I'm just trying to think of a way to say "up yours" without getting fired.
-- Stef
User Friendly, 10/8/1998
pgp00000.pgp
Description: PGP signature
