here is what i sent to a friend i was having this discussion with.
ultimately, vendors like compaq supply a spec sheet on the i/o time per
controller in various configs.  I'd like to get that info for ours.  look
forward to talking about this more - very interesting.

cheers-byron

-----Original Message-----
From: Byron Kennedy [mailto:[EMAIL PROTECTED]]
Sent: Monday, February 04, 2002 12:53 PM
To: NT 2000 Discussions
Subject: RE: SQL Server and RAID Levels


Because when writing to a RAID 5 array (in general), both the target disk
stripe and the target parity stripe must be read(2) and the parity
calculated, and then both stripes written out(2).  As you say, this
performance hit can be vastly improved through the use of write caching,
though the disks still have to do the writes (short-term vs. long term
performance).  Agree with your points though.

Tony, as other have advised, use RAID 1 for the logs.  the i/o to the
transaction logs are almost 100% sequential writes - no perf issues.  you
don't want to loose these in a RAID 0 because 1 drives drops.  if nothing
else use cheaper disks for the logs.

good luck.byron 


-----Original Message-----
From: Wes
Sent: Monday, February 04, 2002 12:30 PM
To: NT 2000 Discussions
Subject: RE: SQL Server and RAID Levels


How does RAID 5 need 4 i/o's per disk?


RAID 5 is slower because it must calculate a parity bit before writing the
data to disk.  If you have enough cache so that the OS thinks it is written
before it is really done then that negates the performance hit of RAID 5.
All controllers we use have 128MB cache and some large SAN systems may have
GB's of cache.  RAID 5 is optimal for reads.  The more spindles the better
by the way.  

If you can afford it RAID 10 or 1+0 combines the benefits of RAID 1 and 0.

I would never recommend using RAID 0 for anything on a server.  Even if it
is just a dev box you still have to consider lost labor hours while people
can't work.

Benchmarking is the only way to tell if you absolutely must have the optimal
cost and performance solution.  We use RAID 5 for all large configurations.
RAID 1 only for the OS and transaction logs.

-----Original Message-----
From: Byron Kennedy [mailto:[EMAIL PROTECTED]] 
Sent: Monday, February 04, 2002 2:21 PM
To: NT 2000 Discussions
Subject: RE: SQL Server and RAID Levels


not exactly:

RAID 5 is the most costly on writes as it requires 4 physical i/o per disk
per write (remember parity is distributed across all drives).  

RAID 1 requires 2 physical i/o per disk, per write. 

RAID 0 requires 1 physical i/o per disk per write.

RAID 10 (0+1) requires 2 physical i/o per disk, per write.

byron

-----Original Message-----
From: Szlucha, Chris [mailto:[EMAIL PROTECTED]]
Sent: Monday, February 04, 2002 12:06 PM
To: NT 2000 Discussions
Subject: RE: SQL Server and RAID Levels


Forgive me if I'm wrong, but Ed wasn't saying to use both RAID 0 and RAID 1,
but rather what is sometimes called RAID 10, or more properly RAID 0+1.   It
is a mirrored RAID setup.  Speed and redundancy, but it's the most costly of
the bunch.

And your statement about it being faster on RAID 0 or 1 is incorrect.  RAID
5 is faster, as the write job is split up across the drives and each drive
writes it's own piece of data at the same time as the others.

-Chris

------
You are subscribed as [EMAIL PROTECTED]
Archives: http://www.swynk.com/sitesearch/search.asp
To unsubscribe send a blank email to [EMAIL PROTECTED]

------
You are subscribed as [email protected]
Archives: http://www.swynk.com/sitesearch/search.asp
To unsubscribe send a blank email to [EMAIL PROTECTED]

Reply via email to