Raw vs. file system
RAW
benefits:
faster (almost all cases)
guarenteed transaction completion by database
completely supported, and "recommended" by all databases
hinderances:
completely rely on app for backups, restores, and access
no expansion without great pain
FS
benefits:
can backup data files with conventional methods
visibility of datafiles without app (obviously not the actual data in
the files)
can expand storage capacity on the fly
moving databases is a matter of changing pointers and moving files
hinderances:
dba's
slower (almost all cases)
no guarenteed transaction completion (generally)
not completely supported by database companies for the most part
Regarding some of the above.....
Ufs is much faster now than it was 3 years ago, some statements of only
10% slower than raw were received. Now 10% may be huge for some, not as
huge for us, plus some pass-through drivers can actually boost
performance to that of raw and (according to Veritas) beyond. Guarenteed
transaction completion can be accomplished on a FS as well, either with
the Veritas route or with a newer ufs. Plus, being able to grab
datafiles off of one machine and copy them to another, then point the
database app to the new location is a gem if your system bites it for
some reason. With FS you can restore a database to a test system that
isn't an exact disk copy of the original.
I am of the opinion that, if you have a budget, you can get everything.
Manageability and versatility of fs with performance and integrity of
raw disk. Most replies stated things I already suspected; that Veritas
FS with Quick I/O can give you better performance than raw disk (if you
have the budget for it), that Solaris has a forcedirectio mount option
that bypasses (or turns off) FS caching and let's the user write direct
to disk (man mount_ufs on Solaris 2.6 and 7, not an option on 2.5.1).
That all database companies "prefer" you use raw disk but have blessed
other types of storage; Oracle runs fine on NetApps and Informix has
blessed NetApp recently as well, Sybase is the only holdout that I know
of.
For NFS it is true that performance will suffer some due to the nfs
protocol, but in my experience for what we do it won't be much.
Some argue that the FS on NetApp also degrades performance (and some
think that ufs is faster for both writes and reads) but it has been my
experience that it is minimal if any. We run all of our design apps
from a netapp (cadence, mentor graphics, vcs, synopsis, etc..) as well
as Netscape browser and mail client for ALL of our Unix clients, nothing
local, no exceptions. Keeps our model manageable. We also store ALL
mail data on a netapp (using Netscape mail server and imap) for everyone
in the company, remote sites included. All design data is stored on
netapps which is read from one and output written to another during
simulations. In a nutshell, our data storage is almost exclusively
netapp with the glaring exception of Informix.
Now, I'm not trying to force a model if it doesn't fit, but noone was
even willing to test it. After the meeting we have come to an
agreement; we go with raw AND netapp, pointers to both, and see what
happens. We'll benchmark performance keeping in mind the limitations
and benefits of each and see where we end up.
My thanks to all who replied, there were many. Had no idea we had so
many closet dba's.....
~JK
Jeff Kennedy wrote:
>
> I am going into a meeting in a few hours concerning this topic. I have
> a couple of dba's that swear we can't go to filesystem over raw but I am
> of the opinion that we can and that the issues of yesteryear have been
> addressed.
>
> We are running Informix on Solaris 2.5.1 - 7. It is all direct
> attached, which is another thing I would like to move away from and put
> this on a NetApp. What I am asking for fromt his group is yea's (and
> why) or nay's (again, why). I'm sure this info is out there in vast
> quantities but time is pressing on me and I am not sure where to begin.
>
> Links to articles are welcome as are experiences and opinions.
>
--
=====================
Jeff Kennedy
Unix Administrator
AMCC
jkennedy@amcc.com
_______________________________________________
codeprof mailing list
codeprof@codeprof.com
http://www.codeprof.com/execute/ask/?codeinfoid=61171
