Maxtor disks & DEC5000
Marcus J. Ranum
mjr at hussar.dco.dec.com
Mon Oct 15 06:51:57 AEST 1990
In article <2933 at canisius.UUCP> pavlov at canisius.UUCP (Greg Pavlov) writes:
>In article <1990Oct11.211023.2207 at decuac.dec.com>, I write:
>> Depending on the way you anticipate using your system, you might
>> want to increase the size of your buffer cache, and change the write
>> scheduling policy to delay scheduling dirty buffers for write.....
>>
> What is the downside of this ? Am I taking a bigger chance on ending up
> with a corrupted database, for instance ? What are potential performance
> downsides ?
The possible downside of the delayed write is a slightly better
chance of losing something. Off the top of my head I can't think of a case
where delaying scheduling dirty blocks for write would be a performance
hit. Unless your machine is really unstable I wouldn't worry about losing
a block or 2 every so often. :)
Bumping the buffer cache can be a performance hit if it causes you
to start running short of memory for jobs you may have running. If you
increase the buffer cache to, say, 20% and notice that you're starting
to swap where you weren't before, you're hurting your performance. If you
have a machine, for example, that is set up as a file server - nobody
logs in and starts running X-windows on it - then boosting the cache
may make a lot of sense. If *all* you're doing is file service, you can
steal a riff from Plan 9 and just use everything you can for cache. :)
> I have played with the tunefs parameters many times - particularly rotat-
> ional delay. I have noticed that I do not see any difference (well, no
> more than 2-5%) regardless of what I set them to and on what type/brand of
> disk (RA70's, RA90's, RDnn's, Fuji M2333k, various CDC and HP).
Yep, I've never managed to squeeze the legendary Massive Benefits
of tweaking rotational delay out of the disks, either. I did manage to make
the disks on my 3100 measurably slower one afternoon, so I suppose I could
say I reaped massive benefits by undoing the damage.
mjr.
--
coffeecoffeecoffeecoffeecoffeecoffeecoffeecoffeecoffeecoffeecoffeecoffeecoffee
More information about the Comp.unix.ultrix
mailing list