[ On Wednesday, December 12, 2001 at 22:44:10 (+0100), Manuel Bouyer wrote: ]
> Subject: Re: RAID-5 benchmark results
>
> On Wed, Dec 12, 2001 at 08:30:56AM -0800, Johnny Lam wrote:
> > Well, who runs production file servers with RAID and doesn't invest in
> > a UPS?  Not me =)
> 
> A UPS doesn't prevent you from UPS failure, or system crashes, or someone
> pluging out the power, or your UPS-monitoring software failing when
> the UPS runs out of battery, or ...
> I've got enouth unclean shutdown on servers protected by UPS to not play
> this game :)

Hang on just a minute here!

Sure there's no triple redundancy here -- but you really don't need it
unless you've got terribly unstable systems and power!  With a UPS on
production systems the risk of damage is nearly nil.  There are so many
if's on top of if's in your caveats that I think only a system with the
most important of data and the least likelyhood of timely offline backup
would ever be at any real risk.  (Obviously this kind of risk assesment
cannot be done by weighing the vulnerabilities alone!  :-)

Now what would be nice would be a little program that could poke at your
disks, RAID, whatever, and cause them to flush their caches in a
predicable way -- just the way we have update, er, ioflush, occasionally
(well every 30 seconds) flushing dirty I/O buffers....  Eg. if there
were a SCSI command I could send to my hardware RAID controllers to
encourage them to write any dirty cache buffers, I wouldn't mind having
a little daemon sit and make such a request every 5-15 minutes or so.

Last night I happened across an old post I made long ago about adjusting
the flush rate to a much longer interval on machines with lots of memory
and slower disks.  At one point I had my 3B2/500 running at a 240 second
interval, which was long enough for it to unpack and store at least two
average sized news batches, thus ensuring that on average very few of
the buffers for temprary files ever had to be flushed to disk.  It
worked like a charm, even without a UPS (since power in this city is
really quite stable and reliable comparatively speaking).

BTW, I think the various *delay variables in miscfs/syncfs/sync_subr.c
should be adjustable with sysctl (though I'm not quite so tempted to
increase them wildly on my current machines, even though I do have a
good reliable UPS! :-).

-- 
								Greg A. Woods

+1 416 218-0098;  <gwoods@acm.org>;  <g.a.woods@ieee.org>;  <woods@robohack.ca>
Planix, Inc. <woods@planix.com>; VE3TCP; Secrets of the Weird <woods@weird.com>