Re: RAID Questions

From: Rich Ashton <[email protected]>
Date: Thu, 9 Apr 1998 09:44:39 +0100 (BST)

On the Raid issue, I run the web cache here at NETCOM UK.

On a Solaris 2.5.1 machine US1, Sparc SSA with 18Gb of disk, 640MB of main
memory I had major problems when running RAID level 5. The parity was
causing the disk busyness, as shown by iostat, to reach 56% when we were
receiving just short of 50,000 request per half hour period. The cache
handled 1.4 million requests per day. The worst part was wait times, upto
457ms at times when disk busyness was high.

I reconfigured the SSA to used a simple striped configuration, spreading the
way it wrote across as many controllers as I could, so I had 9x2Gb disks
across 3 controllers. Squid rarely forces disk busyness above 18% now, and
although I haven't noticed any increase in speed of serving pages, there is
now very little sign of any bottlenecks in the disk I/O, which was a major
problem serving so many users.

I have yet to see how this configuration works on Solaris 2.6, but given the
kernel re-writes I expect an increase in performance (once I've got the code
working that is! Thanks for the patch Mr Wessels!)

BTW - I use seprate partitions for the cache directories, the index files,
and the log files - the index and log files are mirrored using a simple
mirror, since I'd like to be able to repopulate my cache if any of the cache
disks fail. If a cache disk does fail, you loose the whole of the date
(well, almost) since their is no parity for the RAID level I use for the
cache disks - however, speed and cheapness are paramount, the data is
secondary and doesn't need to be mirrored - it takes little time given the
swap file to repopulate the cache (although some would say 2 days is quite a
while!)

Rich.

> I too would love to see a truely useful comparison, but one big
> question comes to mind: What about the drive that Squid writes it's own
> logs (cache.log, access.log, etc.) to? Would this drive hold Squid back if
> the cache was on a RAID array? Would Squid just queue all the log writing
> requests until the drive could catch up? Or would it just be best to leave
> some room on the RAID array to hold Squid's logs?
>
> -Bill

-- 
..Blue         O            "Smoke me a kipper,
  Skies..    //\/            I'll be back for breakfast."
            \/\  ..Must
         ...../    Dash..    Email: rich@ops.netcom.net.uk
Received on Thu Apr 09 1998 - 01:51:08 MDT

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:39:38 MST