Raid 1 disk scheduling

Raid 1 disk scheduling

[David F. Bakker]

I have a server with lots of available RAM (10GB) I
wanted to create an md raid 1 device that would mirror
a large ramdisk (2GB) to a scsi md device. The
thinking here is writes will be just as fast as always
since I/O is sync to both the RAM and hard disk.
However reads should be crazy fast since it should be
coming from the device with the lower latency or the
one that answers first (RAM). It looks like the I/O is
done in a round robin setup. I get amazing reads every
other test (cat'n a file) which seems it is a round
robin scheme. Can this be changed somehow to prefer a
device or use latency instead?


		
__________________________________ 
Do you Yahoo!? 
Yahoo! Small Business - Try our new Resources site
http://smallbusiness.yahoo.com/resources/

Re: Raid 1 disk scheduling

[Molle Bestefich]

David F. Bakker wrote:
> I have a server with lots of available RAM (10GB) I
> wanted to create an md raid 1 device that would mirror
> a large ramdisk (2GB) to a scsi md device. The
> thinking here is writes will be just as fast as always
> since I/O is sync to both the RAM and hard disk.
> However reads should be crazy fast since it should be
> coming from the device with the lower latency or the
> one that answers first (RAM). It looks like the I/O is
> done in a round robin setup. I get amazing reads every
> other test (cat'n a file) which seems it is a round
> robin scheme. Can this be changed somehow to prefer a
> device or use latency instead?

Linux should use any available RAM for disk cache.
Are you sure that using MD as a disk cache like above is any faster than that?

Re: Raid 1 disk scheduling

[David F. Bakker]

--- Molle Bestefich <molle.bestefich@gmail.com> wrote:
> David F. Bakker wrote:
> > I have a server with lots of available RAM (10GB)
> I
> > wanted to create an md raid 1 device that would
> mirror
> > a large ramdisk (2GB) to a scsi md device. The
> > thinking here is writes will be just as fast as
> always
> > since I/O is sync to both the RAM and hard disk.
> > However reads should be crazy fast since it should
> be
> > coming from the device with the lower latency or
> the
> > one that answers first (RAM). It looks like the
> I/O is
> > done in a round robin setup. I get amazing reads
> every
> > other test (cat'n a file) which seems it is a
> round
> > robin scheme. Can this be changed somehow to
> prefer a
> > device or use latency instead?
> 
> Linux should use any available RAM for disk cache.
> Are you sure that using MD as a disk cache like
> above is any faster than that?
> 

No Im not. I would assume it would be since if I start
a large service (1GB) and then stop it and restart it
it takes just as long. I was thinking since everything
would already be in ram that it would start much
faster. However after thinking about this somemore
it's probaly the apps internal init that is slow and
not the actual executable load time. I did make a
change to raid1.c to favor ram disks if present which
shows some great read speeds but I havent tested the
whole thing out yet.


		
__________________________________ 
Yahoo! Mail Mobile 
Take Yahoo! Mail with you! Check email on your mobile phone. 
http://mobile.yahoo.com/learn/mail