Re: OT: Re: [exim] Reducing Disk IO - Speed Issues

Good morning,

Nigel Metheringham ecrit:
> On Thu, 2005-08-25 at 11:10 -0400, Walt Reed wrote:
>>On Thu, Aug 25, 2005 at 03:58:51PM +0100, Nigel Metheringham said:
>>>On Thu, 2005-08-25 at 16:54 +0200, V. T. Mueller wrote:
>>>>@Marc: a battery-backed write cache is an optional hardware feature
>>>>for the better controllers. This is somewhat paradox, because on
>>>>important high-IO systems - which is where you usually use rather
>>>>expensive RAID controllers - you rely completely upon UPS triggered
>>>>shutdown routines in case of a power failure as long as your
>>>>controller cache is not battery backed. It's quite tricky to restore
>>>>a database that had just lost a couple of megabytes of transactions...
>>>You still need to have write integrity down to the disks, since even
>>>with big expensive UPS systems you still find some clown who manages
>>>trigger the EPO (big red button that turns the power off - normally
>>>carefully mounted in places where they can be hit accidentally), or
>>>remove the power cable, or switch off the *wrong* machine.....
>>>Never underestimate the resourcefulness of an idiot in a data centre.
>>Battery backed write cache has NOTHING AT ALL to do with a UPS. The
>>battery is on the RAID controller itself.
> Yes. However vt appeared to be claiming the battery backing was
> unnecessary due to the UPS, I was responding that there are faults the
> UPS doesn't protect against.

Sorry for my unclear wording. I wanted to point out that it is
strange that many people spend a rather high amount of money for a
raid controller /without/ battery backed cache in order to improve
data reliability and access performance. They usually do this
because the usage of the system is important enough to do so.
By doing this, they put a lot at risk and rely completely on UPS
shutdown routines. Without a UPS or properly installed and tested
shutdown routines a tremendous amount of data can be lost in the
event of a power failure. In that case, raid and/or journaling won't
help much.

> Except of course where there is a more interesting failure mode. The
> best example I have seen recently is where the neutral got swapped with
> one of the phases - OK for 1/3 of the customers, the others had a range
> of interesting faults.

Well, an online ups usually detects this and triggers the external
fuse. Nevertheless I'm not stating your data is safe just because
you're properly using decent ups hardware. My point is you rely on
it if you don't use battery backed controllers (and deactivated disk
caches) - adding a spof to your environment. The (additional) bitter
taste of the example above is that those 1/3 of customers are not
automatically on the winning side. They have a high risk of delayed
hardware failure - in wich case no insurance will pay.

Cheers,
vt