Sun A1000 RAID arrays and AFS

Terry McCoy terry@nd.edu
Mon, 7 May 2001 21:50:22 -0500 (EST)


On Mon, 7 May 2001, Mark Montague wrote:

> 
> Hi,
> 
> Is anyone using Sun A1000 RAID arrays with AFS?  If so, what have been
> your experiences?  Good?  Bad?
> 
> We're considering buying some under a special educational deal (the
> price is too good to pass up) but thought we'd check with others first.
> 
> Thanks in advance.
> 
> 

We have 11 AFS file servers using E250 (2 x UltraSPARC-II 400MHz)
256 MB RAM and one A1000.  All of our A1000's have 64MB of cache on
the controller.  We have a mixture of 8 and 12 disk drive A1000
arrays.  The 8 drive arrays have 36GB 10K RPM while the 12 drive
arrays have 18GB 10K RPM drives.  I have configured the arrays with
LUN's using 4 disks with a 64KB stripe.  The entire LUN is made into
one AFS partition.  On some of the E250's we use the internal disks
as an additional AFS partition using Veritas to construct a 0+1 raid.

At one time when space was really tight I had an 8 drive array with
one LUN and like 230GB AFS partition.  I *wouldn't* recommend that as
workable configuration.

I use the small 4 disk LUN's because when you have a drive go bad
the reconstruct of the LUN will be shorter when the bad disk is
replaced.  Also in theory if you had multiple drive failures and
those drives were assigned to different LUN's you would still be
up.

Our configuration is:

    Solaris 2.6  (Generic_105181-21)  patches from the 08/01/2000 patch report
    Sun RAID Manager 6.1.1 Update 2
    SEVM 2.6  (REV=2.5.3)   with the patch 106606-03 and afs3.6 2.3
    AFS file server (Base configuration afs3.6 2.3)


One thing you should reconfigure is the default configuration for Sun
Raid Manager.  You should disable the automatic parity check, run it
from cron maybe twice a month.

I have found that a reconstruct of a 4 disk LUN using the 36GB disks
took about 5 1/2 hours on a file server that had no AFS volumes (hence
no load).  I believe the reconstruction was optimized for reconstruction
performance.


Does anyone else have any data on reconstruction performance?

tHanKs
--
Terry McCoy                             email:  terry@nd.edu
Sr Systems Engineer                     phone:  (219) 631-4274
Office of Information Technologies
University of Notre Dame