This post is more than 5 years old
43 Posts
1
1612
March 18th, 2010 11:00
High stripe crossing with diskpart
Hi there,
I've a problem with a Windows 2003 LUN.
I've created a 75GB LUN in my CLARiiON CX4 with
Element size : 128
Alignment Offset: 0
RAID 5
FC Disks.
I've asigned the LUN to a windows 2003 box, then I've used diskpart for align the partition to 64 - 256 and 1024kb, then i've formated the partition
NTFS with 8k blocks.
I've ran :
# iozone -s 1g -r 8k -t 4 (RAID5) -i 2 (randon reads/writes)
When iozone finalized i saw the LUN statistics and the results are :
Align = 64, blocks = 8k and NTFS 8k
Stripe Crossing = 7319 ( with iozone 8k)
Stripe Crossing = 17200 (with iozone 4k)
Align = 256 , blocks = 8k and NTFS 8k
Stripe Crossing = 10419 (iozone 8k)
Stripe Crossing = 9119 (iozone 4k)
Align = 1024, blocks = 8k and NTFS 8k
Stripe Crossing = 7114 (iozone 8k)
My question it's simple,
It's normal stripe crossing with disk alignment ???
I'm very confused.
Can you help me to understand this?
best regards.


faisalchoudry
53 Posts
0
March 29th, 2010 02:00
Which Flare code are you running ?
Have a read through " EMC Clariion Best Practices for performance and Availability: Release 28.5 Firmware Update". There a similar document for Flare 29 as well.
Look at the alignment section, around page 20, it's a good explanation, and discusses where you will see benefits, not all environments, and it all depends on the operating system, the application, the IO size wriiten, the Raid group and the spindle count, and how much you've aligned by.
Did you also run your tests on the same type of LUN and Raid Group with no alignment set? what were the results ?
You will get crossings, but they should be reduced, by what percentage depends on the I/O size been written.,
There is an easy formula in the FLARE 26 paper which you can use to calculate the approx percentage decrease in crossings, by aligning.
Percentage of data crossings = (I/O size) /(stripe element size) * 100
Larger IO sizes tend to be more effected by crossings.
Some enviroments which I've gone onto, the customer has asked whether I could align their partition because it wasn't originally done, but based on whatever is being written to that particular LUN, it might not be worth the hassle, as there may be no gain, and setting the alignment on an existing LUN in Windows for example...well you can't, you'd have to make a new LUN.
However for a brand new partition and LUN, I can't see any reason not to set the alignment.
The Flare 26 paper had some extra bits explained on the section discussing alignment, not in the Flare 28 or 29 papers (i know you have CX4 but you may find this interesting to read).
If you can't find the Flare 26 paper, let me know and I can email that to you.
kelleg
6 Operator
•
4.5K Posts
1
March 18th, 2010 11:00
In Which order did you run the tests - 1024 first and 64 last?
I believe that stripe crossings are cumulative. Also, how many disks in the raid group?
Glen
igarcia1
43 Posts
0
March 20th, 2010 02:00
Hi kelleg,
I started the test in 64k always with Raid5 (4+1) then 256k and then 1024k.
I've reset the statistics loggins in the SP before each test.
igarcia.-