raid5 - software, hardware, performance, kernel 2.6

Andrew Lentvorski bsder at allcaps.org
Fri Mar 17 23:57:17 PST 2006


Stewart Stremler wrote:

> Poking around at Wikipedia, I get the impression that SATA requires
> one controller per disk (but it's only seven wires, so adding another
> controller to the ASIC is fairly trivial), with no daisy-chaining.

Yes, and one controller and cable per disk is a *good* thing.

For example, a hot swap failure won't take down the entire chain.  A 
terminator failure won't take down the entire chain.  A cable ground 
fault won't fry multiple drives.

> Somewhere I found an assertion that SATA-II will support up to 15
> devices per controller, but it also talked about SATA port replication.

Here is the official site:
http://www.sata-io.org

> How much CPU time does a SATA disk require?  SCSI can do device-to-device
> transfers with comparatively little CPU involvement; that, plus the number
> of devices you could hang off of a controller, makes SCSI much more
> attractive to me than IDE/ATA/PATA.

Try here for starters:
http://www.sata-io.org/docs/SATA%20Enterprise%20WP.pdf

Serial ATA does have command queuing like SCSI.  Drive to drive transfer 
is a function of the controller chip since one drive does not have a 
direct connection to another drive.

Also, serial-attached SCSI seems to be the development path of the 
future.  So, SCSI is going to lose drive to drive.  The enterprise folks 
don't seem to be bothered too much by that.  Personally, I agree with 
them.  I would rather see the drive to drive transfers handled by my 
controller card rather than stuck into the drive.  That way I get the 
best performance my *card* supports rather than being stuck with the 
worst performance of all of my drives.

Letting the controller handle drive to drive is a *BIG* win when doing 
read-modify-write cycles like are required for RAID 5 or 6 parity striping.

> Are these for comparable hardware?  Another one of the web-pages
> asserted that (Enterprise-class? Server-class?) SCSI drives have
> stronger arms and better motors and suchlike, to better handle 
> server-class loads.

That is true.  SCSI drives need stronger motors for 15K RPM drives and 
they up the seek servos to match.

However, some SATA drives are now up to 10K RPM.  The big problem with 
this is the thermal management and power consumption of the faster 
drives.  Desktop class hardware simply can't handle 15K RPM drives.

-a



More information about the KPLUG-List mailing list