Subject: Re: differential scsi controllers
To: None <current-users@NetBSD.ORG>
From: der Mouse <mouse@Rodents.Montreal.QC.CA>
Date: 07/02/1998 10:00:22
> Disks, and the kinds of disk-support environments one tends to find
> in most PC cases today, are not yet reliable enough to where I can
> take (for example) 5 drives and ccd them together without
> *substantially* increasing the chance of catastrophic failure for
> whatever data is on that volume.
And they cannot be; if the chance of failure of a disk is P (for some
reasonable time interval), the chance of a failure of a two-disk ccd is
1-(1-P)^2, or 2P-P^2, which for anything approaching usable failure
rates is only marginally less than 2P...which qualifies as a
"substantial" increase. Regardless of how tiny P actually is.
Now, I don't know how bad peecee hardware is. But this is being
discussed on current-users as well as port-i386 (and indeed I'm
dropping the latter as it really isn't port-i386-specific). And not
all disks and disk housings are bad enough that 2P, or even 4P or 5P,
I work for a lab with roughly 80 gig of disk on its main fileserver
machine (26 spindles), and probably another 10 to 20 gig scattered
around on various other machines - overall, probably 30 to 40 spindles.
I think I can count the number of disks that have failed in the past
year, across the whole site, on the fingers of one hand (indeed, I'm
having a hard time recalling more than one troublesome disk).
> [...] I still don't like the numbers I get when I take the TTBF for a
> single drive and divide it by 5 to estimate that ccd's chances of
> providing uninterrupted service for 12 months.
If I assume I've forgotten two-thirds the disk failures that have
occurred in the past year at work, that means three failures in (say)
36 spindle-years. Stripe five such drives together and you get,
roughly, one failure every couple of years.
That's acceptable to me - heck, I have *CPUs* die more often than that.
Perhaps you have more stringent requirements. Or perhaps you're stuck
with crappy drives (if so, yes, a ccd may not be the way to go).
7D C8 61 52 5D E7 2D 39 4E F1 31 3E E8 B3 27 4B