> pgsql@mohawksoft.com wrote:
>
>>You are absolutely wrong on all accounts here. A RAID5 system is slower
>>than a single spindle as it is only as fast as the slowest disk in the
>>stripe and the overhead of the RAID.
>>
> Huh, what kind of controller do you use... Sounds like some "value" IDE
> one. I'd never suggest IDE raid5 for DBMS purposes anyway.
Actually, my RAID system, currently on my test system, is fully UWLVD SCSI
with fast spindles.
Here is a logical factual question for you to answer: how can a set of
disks, lets say 7, 6 data drives with one parity, deliver results faster
than the slowest drive in the stripe?
If you say predictive and intelligent caching, yea, maybe, but *all* disks
today have caching, but the initial request still has to wait for the
longest seek time across all spindles and the slowest spindle position.
I've been dealing with RAID systems for almost a decade now, and they are
not a magic bullet.
RAID systems are always slower than their compnent disks. This is the
drawback to using them and a fundimental limitation. A single disk will
average 1/2 spindle seek, assuming its initial head placement is random,
and average 1/2 spindle revolution to track, assuming no out of order
sector access. A RAID system has to wait for the slowest disk, thus while
a single disk can average 1/2 seek and rotation, two disks will not. So,
your raid disk access will ALWAYS be slower or as slow as a single disk
access not including the additional RAID processing.
The advantage to a RAID is that a number of smaller disks can look like a
big disk with some redundency. The advantage to a RAID controller is that
the RAID processing and parity generation overhead is done on an external
device. Using a RAID controller that presents a SCSI LUN is great because
you don't need to trust third party drivers. All in all, RAID is a good
idea, but it isn't faster.
As for IDE RAID, IDE RAID is an awesome idea. SCSI disks are just too
expensive. Infortrend has a cool IDE to SCSI or Fibre RAID system that
rocks.
