SUMMARY : cache + RAID 5

From: Gary Menna <G.Menna_at_isu.usyd.edu.au>
Date: Tue, 16 Jun 1998 12:54:55 +1000 (EST)

Question :
Hi all ,
       We are upgrading one of our Alpha boxes .
       We are getting a couple of the 500+ CPUs
       and a HSZ70 , BA370 with 128 Mb of cache .
       I would like to set the disks up as RAID 5 for use
       with (sybase) databases . Will the 128 Mb of write back
       cache offset the penalty of RAID 5 . ?


Thankyou to the following for their exellent replies .
I will certainly set up a meeting between us System Admins ,
DBAs , and application programers to find out what the work
load is . I anticipate most of the dailly access to be reads
with most of the update/writes done at night by batch jobs .
My devices will be LSM raw disks . I've also been told by the DBA
that sybase can "mirror" the devices . Great ! thats just what
I need , another option !! 8-))) .


Answers :

From: "Alan Rollow - Dr. File System's Home for Wayward Inodes."

Under worst case I/O loads, nothing can offset the problems that
RAID-5 has with writes, except cache as large as the underlying
device. But, a larger cache can certainly in many I/O loads.
The worst case is probably constant random writes that are smaller
than the chunk size of the array. Till the cache fills, it can
absorb all the writes, but once full it has to flush older data
to accept new data and the flush will likely suffer the full write
penalty. If writes are clumped so that multiple, logically
contiguous chunks are affected, the presense of the cache allows
some optimization of the writes.

Constant sequential write loads aren't as bad. It has the same
cache flushing problem, but there's more opportunity to optimize
the writes; enough to make it look like RAID-3.

If there is idle time between writes, the controller has the
opportunity to flush the cache, keeping some of the cache clear
for taking new write data. If you suspect that the data won't
be read back immediately, it may help partly disable read
caching. I read that you can do this by setting the max cache
transfer size to 1. If you're writing through the file system
this may be a useful change, since the UCB will be caching the
data. If you're using a raw device for the database, you may
want the cached for database read-modify-write operations.


>From Girish Phadke :

Hello
BA370 gives you 6 SCSI buses with 4 drives each.
So you are likely to use raid set of how many drives.
If only 3 to 4 drives raid set across the buses with 128 MB cache.
You may not see any performance degradation at all for read operation.
do not forget to enable write back cache.

But is it nesscary to have all RAID 5 one set of mirror disk also you can keep
for such read intesive operation if any.



>From Acacio (Casey) Carvalho :

There is a bit of a penalty for using RAID 5, but the HSZ70s with 128MB of
memory are certainly not the bottleneck. We have pretty much the same
environment on a ESA10000 cabinet, and the performance is certainly
excellent.
We use Oracle under ADVfs (raw devices are a bit faster, but the DBAs loose
some management flexibility). I don't know what your cabinetry is, or how many
disks and RAID5s you have, but in our case we allocated three RAID5s (with 6
drives per RAID5, plus 1 spare), and then striped the RAIDs so that the I/O
actually hits all 18 drives at the same time. We also have 2 HSZ70s
(redundant).





Thanking you ,

Gary Menna E-Mail g.menna_at_isu.usyd.edu.au
Information Technology Services Phone +61 2 9351-6360
University of Sydney (G05) Fax +61 2 9351-7711


      #=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#
      # #
      # Quidvis Recte Factum, Quamvis Humile, Praeclarum #
      # #
      # R-R #
      #=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#=#
Received on Tue Jun 16 1998 - 04:56:12 NZST

This archive was generated by hypermail 2.4.0 : Wed Nov 08 2023 - 11:53:37 NZDT