AdvFS vs. UFS

From: Christopher Knorr <cknorr_at_hops.com>
Date: Thu, 22 Oct 1998 10:46:05 -0400

Our company has a database product whose primary purpose in life is to
perform very fast ad-hoc queries against very large databases. Because of
our 'need for speed' FAST I/O is CRITICAL. To give a bit more technical
detail about our database we do some fancy stuff with heuristics and file
compression and do our own file striping across multiple logical volumes,
typically configured as RAID5.

>From the beginning we've always used UFS for our data disks. Our reasoning
was we wanted to preserve our interoperability, but because we've stuck to
DU this hasn't really been an issue. From the beginning Digital has
*hounded* us to switch to advFS. Consultants, sales reps, support folks, it
didn't matter. If you even mentioned you were using UFS to anyone a frown
came to their brow, they shook their head sadly, and then started in on why
this was a very bad idea. Because PERFORMANCE is our criterion we thought
this would impact the argument, but NO! AdvFS is FASTER. AdvFS is SAFER.
AdvFS is BETTER.

Okay, fine. I've just completed a rather exhaustive series of benchmarks.
My RAID controller was a SWXCR card. I tried JBOD, RAID5. I modified the
block size of the controller (default is 8; cranked it up to 64). I
modified several advFS tuning parameters suggested in this list. The bottom
line is under NO circumstances did I ever see advFS outperform UFS. The
only way the comparison was even close was to crank up the default block
size to 64, but in order for our database to take advantage of this we had
to set aside considerably more memory. My benchmarks included both a "raw
I/O" test (database independent) and queries using our product.

Well I could live with a little bit slow, but overall our database runs ~
20% slower using advFS. Perhaps worth the tradeoff for some customers I
think, because at least advFS is SAFER. Except I read a day or so ago in
this list where somebody blew out a RAID5 stripeset because a disk failed
and Digital support sez that advFS doesn't handle hardware failures too
well. (ouch!)

So, ... the question that begs asking is, .... Why Does Anyone Use advFS
for their data warehouses? Sure, avoiding the occasional fsck is nice, but
from where I sit the disadvantages seem rather steep. Any thoughts
appreciated - I'll be glad to share any of the technical details/benchmark
results for anyone interested. (i.e. tuning parameters tried, etc.)

Chris

 
Received on Thu Oct 22 1998 - 14:50:57 NZDT

This archive was generated by hypermail 2.4.0 : Wed Nov 08 2023 - 11:53:38 NZDT