RAID on ICHxR

[8steve8]

http://www.alternativerecursion.info/?p=31

bump/comment/digg if you like

 

[supremelaw]

Excellent research: MANY THANKS!

I plan to read it in detail later.

For now, here are just a few data points from
Hard Knocks University:

We compared ICH7R on an ASUS P5WD2 Premium (i955X chipset)
with a Highpoint RocketRAID 2322 installed in an x16 mechanical slot.

ICH7R was wired to 4 x WD 250GB SATA/3G HDDs

2322 was wired to 8 x WD 160GB SATA/3G HDDs.

We were striving for maximum speed while updating
a 90,000 file HTML database, using XCOPY e.g.

E:\> xcopy supremelaw.org x:\supremelaw.org /s/e/v/d

Where,

x = HDD networked with XP/Pro x32


We had to custom-wire the ICH7R to an external chassis,
and that wiring may have been the cause of the problems
we observed: about 3 times, I had to rebuild that RAID 0
because one of the HDDs appeared to "drop out" --
even though the suspect drive was actually AOK
upon closer examination.

This "drop out" happened, even after we installed the
latest ICH7R device driver.

Because I know one of Intel's engineers who developed
NCQ for Intel's Matrix Storage Technology, I do really
doubt that it was a device driver error:

my current theory is that our custom cabling was just too much (too cheap?)
for the voltage assigned to the motherboard's ICH7R ports:

in a word, I was trying to use those ports as eSATA ports,
but the spec for eSATA ports requires a higher line voltage,
reportedly.

The 2322, on the other hand, comes with cables that
plug into the controller at the rear panel. We had
absolutely no problems with this RAID controller
or with any of the WD 160GB HDDs wired to it
in this fashion.

Those 8 HDDs have since been moved into a
larger Thermaltake Armor chassis, and
are now controlled by a new RocketRAID 2340:
this wiring is all internal -- straight from the
controller to those 8 HDDs. So far, so good.


Thanks for the G-R-E-A-T research!


Sincerely yours,
/s/ Paul Andrew Mitchell, Inventor and
Webmaster, Supreme Law Library

All Rights Reserved without Prejudice

 

[mooseracing]

Suprisingly to me, at least from everyones comments on how good the 640's were was the bandwidth they have. Everyone says how outdated the Raptors are, but I'm only seeing about 10MB/s at the most over the min and avg. The Raptors have better burst and access on reads that is.

Interesting data though.

 

[supremelaw]

Another variable that needs experimental control is NCQ -- ON or OFF.

Because it requires a certain amount of computation when enabled,
in streaming situations where lots of sequential data are written then
read back without other I/O requests competing for the R/W armature,
NCQ actually adds a measurable amount of computational overhead --
which can be eliminated simply by disabling NCQ.

Depends on the application (as always): Your mileage may vary.


Sincerely yours,
/s/ Paul Andrew Mitchell, Inventor and
Webmaster, Supreme Law Library

All Rights Reserved without Prejudice

 

[LostInLine]

Here is a discussion about Raid5 on nV chipsets. They do note that the intel raid is much better than the nV raid. But, some things the talk about MAY impact ich9/10r performance. These would be using 3 or 5 disks for raid5 to make the stripe size add up to the cluster size and setting the first sector to align to a cluster and strip width boundary (which cannot be done with XP). See the first post its pretty good.

These techniques may not have an impact on sequential but could have a positive impact on random writes.

[edit: added link]