Re: Measuring IOPS and Raid penalty

How many users? What kind of users? What does your environment look like?

I've been in many environments, and my experience is that the MS assumptions
for light, medium and heavy users are low. I always measure. My
experience has been that the average IOPS/user is closer to 1.5. Things
that impact that number:

Mapi applications (Blackberry, Goodlink, Unity, Rightfax, etc.)
Desktop search, email client search applications (Google, Lookout, etc.)
Average mailbox size
Average number of items per folder
Attachment size limits

I have seen many environments where there are no limits, voicemail and fax
are integrated with exchange, all users have a crackberry, and the standard
desktop deploy image contains a desktop search engine. In these
environments, the measured IOPS/user has been in excess of 3. Clearly the
assumption of 1 IOP/heavy user is nowhere close. You combine that with RAID
5 and an IOPS/spindle assumption which does not account for the IO response
time, and you have an underperforming mess.

Clearly, RAID 10 will give you better performance than RAID 5 for the same
type and number of spindles. Moving to 15K RAID 10, with a 657 requirement
and a 2:1 reade/write ratio means you will need enough spindles to reach
867.14 or 8 spindles in your RAID 10 set. To reach the measured 657
requirement with RAID 5 and a 2:1 read/write ratio you would need 13
spindles in your RAID 5 set. That's a 38% reduction in spindle count using
RAID 10 vs. RAID 5.













"SW" <SW@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote in message
news:0A622870-6EE9-4DBD-8CB0-7D4564426B7A@xxxxxxxxxxxxxxxx
Also is 657 high?

"SW" wrote:

John, based on my info I have given you is this current IOPS too high.
It's
a RAID 5 with 6 10,000 ROM disks. We will go to a RAID 10 with 15,000
disks
in hope of seeing an improvement, what do you think?

"John Fullbright" wrote:

750*.73 = 547.5

Add 20%

547.5*1.2 = 657 IOPS for Database

Now calculate the logs. Next, calculate your read/write ratio and
apply the
write penalty for your RAID type.

For Example:

Database disk requirement with a measured read/write ratio of 2:1.

Reads =657*.66 = 433.62
Writes = 657*.33 = 216.81

The write penalty for RAID 10 is 2:

Required IOPS capacity = 433.62+(216.81*2) = 867.14

That's about 10 spindles at 85 IOPS/spindle or 8 spindles at 120
IOPS/spindle

The write penalty for RAID 5 is 4:

Required IOPS capacity = 433.61+(216.81*4) = 1300.86

For raid 5, first we figure out the number of data spindles -1300.86/85
=
15.3 for 10K SCSI or 1300.86/120 = 10.845 for 15K SCSI. Then you need
to
add in parity spindles. 1/6 on HP or up to 1/11 on EMC.

RAID 5 on HP with 10K spindles = 19
RAID 10 on HP with 10K spindles = 10
RAID 5 on HP with 15K spindles = 13
RAID 10 on HP with 15K spindles = 8
RAID 5 on EMC with 10K spindles 18
RAID 10 on EMC with 10K spindles = 10
RAID 5 on EMC with 15 K spindles = 13
RAID 10 on EMC with 15K spindles = 8 spindles

Of course for NetApp ther is no write penalty:

RAID 4 10K = 9 spindles
RAID 4 15K = 7 spindles
RAID DP 10K = 10 spindles
RAID DP 15K = 8 spindles

John





=


"SW" <SW@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote in message
news:A3D7EDF6-E941-4E8C-AA14-CE0E0A26E5FC@xxxxxxxxxxxxxxxx
If I use our peak then that is 0.730:

IOPS/mailbox = (average disk transfer/sec) ÷ (number of mailboxes)

our average disk transfer/sec peak was 0.730 and the number of
mailboxes
are
750

Would this equal 730/750 = 097 or 0.730/750 = 9.7



"John Fullbright" wrote:

"IOPS/mailbox = (average disk transfer/sec) ÷ (number of mailboxes)"

According to "Optimizing Storage Performance for Exchange Server
2003"
you
should be using peak, not average. Think about it. If you use the
average,
you'll be undersized and performing poorly 50% of the time. If my
minmum
is
1, my average is 5, and my peak is 10, then if I design for average
...
This is a common sizing mistake.

From the paper

"8.
Identify the ex2003base server that experienced the highest
load.
Use
the data collected from the server with the highest load as your
server/processor/storage baseline. Use the following best practices:

. Always design your system to allow 20 percent more
utilization
than you expect for peaks. This allows the storage and processors to
handle
spikes during peak periods.

. Megacycles per mailbox and IOPS per mailbox change as
the
server configuration changes. The following list includes potential
factors
that can change the given megacycles per mailbox and IOPS per
mailbox.

. Mailbox sizes are changed significantly

. Max message size is changed significantly

. Third party applications are added or removed

. Exchange features are added or removed.

. Average concurrency of the users changes (more or less
users
are online using the system at any given time).


9.
After the spread*** (included with the download of the guide
Optimizing Storage for Exchange Server 2003) is populated and your
mailbox
profiles are determined, you can design your storage solution. For
example,
if your analysis indicates that your standard mailbox profile
translates
to
..75 IOPS per mailbox and 1.25 megacycles per mailbox, you can
determine
the
following requirements for a 4,000 mailbox server:

. Mailbox Count: 4,000

. Peak DB IOPS: (4,000 × .70) = 3,000

. Peak Log IOPS: (DB IOPS/10) = 300

. Peak megacycles: (4,000 × 1.25) = 5,000 megacycles


To handle spikes, you should add a 20 percent buffer to your
processor
and storage design. With the addition of this buffer, the minimum
hardware
requirements for this example are:

. Mailbox count: 4,000

. Peak DB IOPS: (3,000 + 20%) = 3,600

. Peak Log IOPS: (300 + 20%) = 360

. Peak Megacycles: (5,000 + 20%) = 6,000 megacycles

"



The standard in the same referenced paper is average write latency
less
than
20ms with no peaks lasting more than a few seconds over 50ms. You
should
size for a 20ms IO response time. A lot of people just sum the
average
sekk
time and rotational latency of a disk and divide one second by that
to
determine how many IOs a spindle can sustain. This method does not
take
into account IO response time. For example, lets take a 15K RPM
SCSI
spindle. For the following IOPS/spindle numbers, here are the
response
times you can expect:

10ms 87 IOPS
20ms 125 IOPS
50ms 200 IOPS

Not taking into account required response times is probably the
second
most
common sizing mistake.








"SW" <SW@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote in message
news:FA50BED9-8B11-4118-AD1E-348D6BC0E8E0@xxxxxxxxxxxxxxxx
We have Exchange 2003 with a RAID 5 volume which has 6 x 10,000
SCSI
disks.

We are trying to work out if we move to RAID 10 have 15,000 would
benefit
us.

To measure IOPS/mailbox

We Used the System Monitor tool to monitor Physical Disk\Disk
Transfers/sec
counter over the peak 2 hours of server activity.

To calculate our current IOPS/mailbox we used the following
formula:

IOPS/mailbox = (average disk transfer/sec) ÷ (number of mailboxes)

our average disk transfer/sec was 0.193 and the number of
mailboxes are
750

Ours

193.531/750= 0.258041 What does 0.258041 mean? We need to then
go
further
and compare our current RAID 5 speed etc with the RAID 10
configuration
we
want to put in.

PLEASE can you explain you calculations to us to so we understand
all
of
it?

Many thanks in advance!









.