Linux Software RAID5 Performance
Velocet
math at velocet.ca
Wed Apr 10 23:07:43 PDT 2002
On Wed, Apr 03, 2002 at 02:49:00PM -0500, Michael Prinkey's all...
> Indeed, the multiple processes accessing the device made significantly
> degrade performance. Fortunately for us, as well, access speed is limited
> by the NFS/SMB and the network, not by array performance. Unfortunately,
> the unit is online now and I can't fiddle around with the settings and test
> it further.
>
> WRT reliability, we have seen the array drop to degraded mode because of a
> single drive failure. We have also a single drive take down the entire IDE
> port. This results in the md device disappearing until you swap out the
> offending drive and restart the array. There is no data here. Usually one
> drive goes and the array goes into degraded mode and starts reconstructing
> on the spare. Then the second goes and the array disappears. It is a bit
> disconcerting to do ls /raid and get nothing back. Changing out the drive
> and restarting pulls everything back.
>
> I can honestly say that the only data loss that I have had on these arrays
> came when a maintenance person completely unplugged one of the arrays from
> the UPS. It caused low-level corruption on 5 of the 9 drives in the array.
> We ended up using a Windows 98 boot floppy with Maxtor's Powermax utility to
> patch them all back up. It took many hours. This is the WORST possible
> scenario, BTW. Even reseting the system gives the EIDE devices a chance to
> flush their caches and maintain low-level integrity. Cutting the power can
> leave the array/drives inconsistent on the filesystem, device (/dev/md0),
> and hardware-format datagram levels. So, lock your arrays in a cabinet! 8)
ok get an EIDE RAID controller with battery backed-up ram onboard. We pulled a
bunch of the SCSI equiv of such from a netfinity server a customer pawned off
on us. Rather nice. (anyone want to buy? :)
/kc
>
> Mike
>
> >From: Jurgen Botz <jurgen at botz.org>
> >To: mprinkey at aeolusresearch.com (Michael Prinkey)
> >CC: beowulf at beowulf.org
> >Subject: Re: Linux Software RAID5 Performance
> >Date: Wed, 03 Apr 2002 10:25:31 -0800
> >
> >Michael Prinkey wrote:
> > > Again, performance (see below) is remarkably good, especially
> >considering
> > > all of the strikes against this configuration: EIDE instead of SCSI,
> >UDMA66
> > > instead of 100/133, 5400-RPM instead of 7200-RPM, and master/slave
> >drives on
> > > each port instead of a single drive per port.
> >
> >With regard to the master/slave config... I note that your performance
> >test is a single reader/writer... in this config with RAID5 I would
> >expect the performance to be quite good even with 2 drives per IDE
> >controller. But if you have several processes doing disk I/O
> >simultaneously you should see a rather more precipitous drop in
> >performance than you would with a single drive per IDE controller.
> >I'm working on testing a very similar config right now and that's
> >one of my findings (which I had expected) but our application for this
> >is not very performance sensitive so it's not a big deal.
> >
> >A more important issue for me is reliability, and I'm somewhat
> >concerned about failure modes. For example, can an IDE drive fail
> >in such a way that if will disable the controller or the other
> >drive on the same controller? If so, that would seriously limit
> >the usefulness of RAID5 in this config. In general how good is
> >Linux software RAID's failure handling? Etc.
> >
> >:j
> >
> >
> >--
> >Jürgen Botz | While differing widely in the various
> >jurgen at botz.org | little bits we know, in our infinite
> > | ignorance we are all equal. -Karl
> >Popper
> >
> >
> >_______________________________________________
> >Beowulf mailing list, Beowulf at beowulf.org
> >To change your subscription (digest mode or unsubscribe) visit
> >http://www.beowulf.org/mailman/listinfo/beowulf
> >
>
>
> _________________________________________________________________
> MSN Photos is the easiest way to share and print your photos:
> http://photos.msn.com/support/worldwide.aspx
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
--
Ken Chase, math at velocet.ca * Velocet Communications Inc. * Toronto, CANADA
More information about the Beowulf
mailing list