Poor XP256 sequential write performance

[Ikuo Fukuda]

I compair sequential write performance between XP256 and DF400( Hitachi's Cheap and small disk array) .
I run "#prealloc 1000000" on both system.
XP256 take 100 sec ( 10Mb /Sec)
DF400 take 35sec(25Mb/Sec)
Both system use only single array group.
Why expensively xp256(with 4Gb cache) have such poor performance?

[Dave Wherry]

I just tried this on my system and saw very differenct results. On my V2500 I have Jamaica (JBOD) disks and an XP256 with 8GB of cache. The results of "timex prealloc junk 1000000" are:
....Jamaica....XP256
real 0.24......0.11
user 0.00......0.00
sys 0.04......0.04
The Jamaica is SCSI and the XP is FC-AL.

The XP was 54% faster. I doubt the problem you are experiencing is inherently related to the XP. There is something else in your configuration.

[Victor BERRIDGE]

Well,
You are on scsi with your XP256 (or HDS7700E) and on your DF400 (HDS58XX or newer model, your results should show the DF400 faster than the XP256 (by up to 50% depending what, I have HDS internal facts that confirm...) these subsystems were developped for the best performance possible, but hasnt the security and comfort of an HDS7700E or XP256:
I have one 5800 connected to an 4 node SP2, 2 time 2 with HACMP, you cant imagine the pain it is when I have to extend the subsystem, and so having to explain to different departments that I have to shutdown all nodes connected in order to add an array for a node that is not theirs, although they have paid for High availability...
We have another one on a another site, but there is no remote copy facility etc...

All the best
Victor

[Ikuo Fukuda]

Thank you Dave.
It is good idea to compair JBOD and xp256.
Sorry I had mistype prealloc command parm on this forum.
I allocated 1Gb ( Not 1Mb)
I also did compaire JBOD disk and XP256.
allocate 1Gb to target disk.

# timex prealloc test 1000000000

JBOD

real 1:14:08
user 0.15
sys 6.55

XP256

real 1:39.80
user 0.20
sys 14.37

DF400

real 38.69
user 0.29
sys 9.95

This result mean that xp256 with 4Gb mem slower than JBOD.
Is it real performance of xp256 with 4Gb mem?

[Dave Wherry]

I do not have any experience with the DF400 so it sounds as if Victor's answer makes the most sense. The DF400 is just faster.
I reran the test here using 1GB (I made the same typo you did) and it looks to me like your JBOD numbers are pretty good, faster than mine. However my XP numbers are much better than yours. Victor mentioned that he thought you were using SCSI on your XP. Is that true? It might explain the numbers.
My JBOD are SCSI and the XP is FC-AL going to that big cache so it better be much faster. My new numbers are:
JBOD
real 2:56:90
user 0.20
sys 14.23

XP256
real 0:40.30
user 0.20
sys 19.90

I'm not sure how much this helps you. At least it's some other numbers to compare with. Like I said, I think Victor really had the answer.

[Ikuo Fukuda]

Our XP256 have 2CHIP(8 FC-AL channel).
Every FC-AL connect to primary and secondary v2500s directly.
So Interface bottle neck was not considered.

Disk of N4000 is just internal system disk.
-----------------------------
My guess is

Huum, Our xp256 that installed last September , its internal disk may be using old 20Mb scsi interface. (@_@)
And Older and cheap DF400s may be use LVD 80Mb scsi as internal interface.

Disks connected to 1 array group may be use 2 old scsi interfaces.
-------------------------------------
Why Daves xp256 was faster than mine?

[Mark van Silfhout]

Hi,

There can be numerous reasons why Dave's XP was faster, like: the number of arraygroups installed, are the arraygroups linear or dispersed formatted, the I/O load on the XP from other systems, the number of ACP's installed, is Continuous Access involved...
Allthough the timedifference is significant the above could cause the difference.

Regards,

Mark

[Ikuo Fukuda]

Thanks Mark.
Ok, the performance of 1 array goup is
limited to about 10mb/sec.
There are noway to get more performance of 1 array group.
If I need more performance, backup huge DB and change array configration and stripe it or
buy more array groups and acp and then use disk striping or buy other faster disk array.

Regards,