On Mon, Nov 22, 1999 at 09:45:52AM +0000, John P . Looney wrote:
> On Fri, Nov 19, 1999 at 03:18:37PM -0800, John White mentioned:
> > Slowing down the writes to the queue is a bad thing. The queue is
> > constantly being sync'ed.
>
> But I thought that messages were all stored in separate files/directories.
> With striping, we can get those spread over a load of disks.
I know. That's why I reccomended RAID 1+0 with writeback cache for sites
which need queue performance.
> > > If he's using Disksuite, he's getting it free with Solaris.
> > Actually, DS doesn't come "free" unless you buy the Server edition
> > of Solaris, at a cost differential to the WS edition.
>
> And if he's running it as a mail server ... Sun are pretty lenient with
> their workstation/server idea. It's supposed to be at odds with microsoft.
> "Here is your $20,000 workstation. We'll be nice and let you use the Server
> CD for free !"
You don't understand. If you purchase a WS edition of Solaris, you don't
get Disksuite.
> > > Hmm. Tell that to Dell. We'ed a few Sun E10000s, some with 4TB+ of disks
> > > hanging off them, one domain alone with 1600 live users, all using software
> > > RAID. Those babies had to have 100% availibility, and a new CPU cost
> > > $16000, so if there were real gains from hardware RAID, we would have used
> > > it. There wasn't any. It was cheaper and faster.
> > For a RAID 1 qmail queue? Or some other RAID topology for some other
> > function?
>
> All the Sun boxes they have run on software RAID - everything from the
> massive pan-European sales & support databases to the little 12 CPU credit
> card authorisation system.
To clarify, my question was:
"Were any of the boxes you mention using software RAID 1 for a qmail
queue?"
Not:
"Please mention some applications which these boxes supported."
Waxing nostalgic about software raid in the general case doesn't make
software raid 1 a good choice for /var/qmail/queue. It's a terrible
choice.
>>> Hardware raid is only of
>>> great benefit for PCs that run NT, or mainframes whose real cost is mips,
>>> so they want to move all processing off the processor. Or people that think
>>> RAID 5 is a good idea :)
>> Not at all. HW RAID is a good way to cache writes, and maximize
>> CPU cycles.
> How does it cache the writes ? Is this CPU cache, or disk cache. With
> software RAID, aren't you just caching anything you write to the disk array
> anyway ?
I don't really mean to be a hardass here, but you need to know about
how the qmail queue works. You have the qmail source right? Included
with that source is an "INTERNALS" document which describes how the
queue works. With qmail's insistance on fsync'ing, you can see how
a writeback cache on the HW RAID controller can help.
Or perhaps you don't know? HW RAID controllers can come with non-volitile
RAM caches. When part of this cache is in "writeback" mode, scsi write
commands are put in the cache, and the controller tells the OS that the
command has been completed. Then the writes are committed to hard drive
(which have their own caches). Thus, multiple small-block writes followed
by fsync's should finish much quicker on a HW RAID with writeback cache.
If you're relying on OS RAM to do the same thing for a filesystem, then
the fsync will put an end to that.
> > It sounds to me like you're abstracting benefits you've seen on software
> > RAID to the qmail queue in a RAID 1. That's not a good thing to do, as
> > qmail has a pretty unique I/O signature.
>
> More than likely :) No, I was more talking about using RAID1+0, so he'd
> get as much benefit from straight striping as possible.
You mentioned that RAID 1+0 would give the max performance, but advocated
software -mirroring- as ok for the qmail queue. It isn't.
I agree that RAID 1+0 or 10 gives the best protected performance.
John
