XP1024 stress performance

[Gordon_3]

Hi all

Our company has just deploy the XP1024 array, since we need to do furture disk capacity planning at the same time, so I just wanna know anybody can give some figures for XP1024, mainly on I/O ( e.g max I/O that each RAID group ( our is 4D + 4D of 36GB disks ) can burst ... ) thx thx.

Gordon

[Leo Simon]

For XP1024 it's very difficult to calculate just one I/O group.
There are several items that might increase or lower your performance such as:
1.The sum of cache and share memory.
2.The raid you had chosen like Raid 0/1 or Raid 5
3.The FC interface card you used and configure
to assign in each RAID group.
4.The ACP card and configuration for each group

So my suggestion for this stress performance you can discuss further this issue with HP consultant and they will suggest you also to
purchase performance monitoring software to see
the real performance in your XP1024

Cheers

[Brian M Rawlings]

Hi. Due to the nature of the XP/Hitachi arrays of this generation, you will find that the performance scales with the size/number of drives, along with other factors (ACP pairs, cache, etc). Unless your I/O is more than 10% writes, or is very sequential in nature, your cache size will not play as significant role as the number of spindles.

This means that you can't really predict overall performance by just testing one raid group, or even using all your raid groups using some benchmarking SW, because overall performance will grow as you add drives. The internal connections in this array are not stressed at all at your current size, so any current throttle or bottleneck is your disks, or the small number of them.

What I'm trying to say is that any numbers you derive from either answers to this question, or from some sort of stress-testing will not be valid except for your current configuration. This is generally good, since you can get more performance (almost linearly) for a long time simply by adding drives and arranging your VGs/LVs to use them all appropriately (striping your LVs across more and more spindles, etc). Not until you get to 60%-75% of the max spindle count for an ACP pair do you start to see busy internal array buses as performance pinch points.

Maybe I'm misunderstanding how many drives you have, but it sounds small for this big box. And, once you add lots more drives, stress-testing is even harder, since few single servers can drive enough I/O to stress this very high performing array (superdomes or the biggest boxes from Sun or IBM or Compaq/Alpha are really the only ones that can do it). Stress testing using multiple boxes is possible, but exponentially more difficult.

Anyway, good luck with the box. Beat it up good, add more disks if you need more I/Os, grow your biz, repeat as necessary.

Regards, --bmr

[Gordon_3]

Hi 2,

Many thx for yr professional explanation, I totally agree with you 2 guys points. The reason I ask this is for our future capacity planning, like how much should I set as threshold level ( e.g. I/O count, Cache utilisation ) indicating the box is nearly overload? ( Actually it's my boss very concern and interest to know .. ^_^ )
Let's share our config. Our XP1024 is now with below config
1. 1 Superdome + 4 N-class + 2 RP class + 4 Sun Micro (E-class) machine attaching to this XP1024
2. 12GB cache with 48 Raid group of ( 4D + 4D ), total 176 X 36G drive

After production, found the current box average I/O is around 7000 - 8000 I/O with max grow up to 9000 I/O. I agree that it's hard to guess how much can this box still "handle", coz from PA, not all the RAID group is "evenly busy", some are quite busy grow up to 900 - 1000 I/O...

So back to my Q, any suggestion on how can we plan better ahead before we really hit the disk bottleneck in this complicated case?

Thx.

Gordon

[Leo Simon]

Hi Gordon,

You can try to use Performance Advisory XP1024
for temporary license and learn about it.
The manual you can download from this site:
http://www.hp.com/cposupport/manindex/hpsurestor64878_eng_man.html
For temporary license you can as HP engineer.

Cheers