Hi,
The original question was basically: Are there any obvious problems
(performance or otherwise) with using LSM to create a volume across two
3-channel RAID controllers, with each RAID controller doing 0+1.
The aim was to achieve greater efficiency (more transfers completed per second)
by having a higher number of spindles.
Thanks very much to the people who replied: Alan, Gunther Feuereisen, and
Danielle Georgette. All of these people provided useful information in a very
timely manner.
The general response was that this setup would be fine. It should give better
efficiency because the load is split over two controllers. Gunther and Alan
gave several other suggestions for increasing this efficiency, such as
making sure write-back cache is enabled, increasing the cache size, and
tuning parameters such as stripe size. It was also suggested that the
RAID controllers be placed on different PCI backplanes if possible, but
our 1200 only has one backplane.
Gunther said that LSM straight off the 4.0D CD with PK3 is fine, and that
no other patches are required.
The original question is attached below.
Scott
> Hi,
>
> On one of our production email servers, we have a filesystem called /spool
> (/var/spool is a symbolic link to /spool). /spool sits on top of a 3-spindle
> RAID 0+1 set (re9 - the second RAID set on one of the RAID controllers).
>
> After looking at the output of iostat over a period of time, we think that
> waiting for IO requests for /spool to complete might be slowing the amount
> of work the system could otherwise be doing. The CPU is about 68% idle, while
> iostat is reporting figures around 120 transfers per second, and 4000kb
> transferred per second. It has been higher at times, but they are the average
> figures. I find it hard to know for sure
> if /spool is a bottleneck for the system or not, just by looking at iostat's
> output.
>
> In any case, our email administrators have requested that /spool be
> reconfigured as two separate two-spindle RAID 0+1 volumes (one on each
> RAID controller), joined into a single logical volume with LSM (ie. mirroring
> and striping done by the RAID controllers, just the joining of the two into
> a single volume done by LSM).
>
> What I'd like to know is, are there any obvious
> problems with doing this? Is it likely to give a significant reduction to the
> average seek time/time to complete each IO request?
>
> Also, do we need to apply any additional patches, or is LSM straight off the
> DU4.0D CD plus PK3 fine?
>
> Additional info:
>
> Operating System: Digital Unix 4.0D PK3
> Machine type: AlphaServer 1200 5/533
> Firmware: 5.2
> Memory: 512Mb
> Disk subsystem: 2 * 3 channel RAID controller (I'm not sure of the type)
> controlling disks across 3 shelves.
>
> I'm happy to provide additional information, and I'll summarise any replies.
> My email address is s.brewster_at_qut.edu.au.
>
> Scott
Received on Fri Sep 10 1999 - 02:01:46 NZST