Re: [SAGE] NetApp--spindles vs. performance

[Chuck Yerkes <chuck@snew.com>]

Quoting Guy B. Purcell (guy@extragalactic.net):
> He's specing out a NetApp filer (F810, I believe), and would like to 
> know if anyone has any real experience with (or even well-educated 
> guesses about) the performance gain of using 36 GB disks instead of 72 
> GB disks.  The intended use will be about 50/50 for typical PC 
> office-type files & Oracle DBs.
> 
> He needs to balance maximizing performance against minimizing cost.  
> Using 36 GB disks costs more (need twice as many disks to achieve the 1 
> TB storage goal), but that might be offset by significant performance 
> gains from having twice as many spindles.  Are there rules of thumb for 
> this sort of thing in general?  There must be some point of diminishing 
> returns with shrinking disk size to increase spindle count (would that 
> I had a system to experiment with!).  Thanks!

"it depends" ....?

First note that for "typical PC office-type" use, you may be
fine with ENORMOUS disks.  And you can have ranks of disk
that are exported for that that are large.

For oracle and DBs in general, I'd offer:
   Why put it on remote storage?  For the pain of network speed access?
We'll get back to that..


I've done work with excessively high scale, high performance
mail work.  We've speced out that we'll use only for first
4-8 GB of any disk in there.  We could only get so much speed
per spindle and we needed maximum speed.

I've also done the same work for a less high perf site.
Same software but far fewer concurrent users, far more
idle mail.  So we let them use the whole disk.

Smaller disks will be faster, but no faster than a larger disk
that you format to use less.  15KRPM disks are notably faster
than 10K and 7200K RPM disks.

Web services can often be massively sped up with caching from
ends (no need to pull logo images across the net for each request).


You only get so much per disk and, with WAFFLE-fs (or any FS),
and so much for a big wad of disks.  The CPU utilization
of NetApps often show when the machine is overloaded.


I'll also presume that you have multiple GB nets bonded to
the Switch to take advantage of this speed.  You don't want
the LAN to be the bottleneck.

I love Netapps, but they are network attached and 1GB ethernet
will give you, in reality, no more than relatively slow direct
attached RAID.

Databases can hammer disks.  My S.O. works with hard hitting
DBs and she has multiple high performance RAID boxes striped
in software across multiple controllers on multiple PCI busses
(first perf.  improvement was taking 3 SCSI/80 controllers
off of the same bus of an E4500).

She's able to get 200+ MB/s (big B) pretty easily and relatively
cheaply.  Gigabit (small B) drivers for the OS's I've seen *might*
give you 60MB/s of actual data throughput.  You'll get far better
DB performance from locally attached disk.


Netapps are a great tool, but not every problem is a nail.