[ale] XFS on Linux - Is it ready for prime time?

[thehead at patshead.com (Pat Regan)]

On 04/23/2010 10:43 AM, Jim Kinney wrote:
> Not having to calculate the checksum for a write is a small speed boost.
> But for R5, the loss of two drives is total loss of data. That checksum
> calculation is a major slowdown on recovery. During recovery performance
> is  order of magnitude slower.
>

The checksums are, or at least should be, irrelevant.  I don't own a 
machine that calculates RAID 5 checksums at speeds lower than a few 
gigabytes per second.

I don't run RAID 5 or 6 anywhere important because of the write 
performance problems.  I do have a server here at home running a 6 SATA 
disk md RAID 5 here at home, though.  It rebuilds just a little slower 
than write speed of a single drive.  I wouldn't be able to rebuild a 
mirror much faster than that.

Is everyone that is getting poor rebuild times on their RAID 5/6 arrays 
using hardware controllers?  Are those hardware controllers using slow 
processors to calculate the checksums?

I don't know what sort of processors are being used on RAID controllers 
today.  I do remember that way back when 1 ghz servers were considered 
fast that high end RAID controllers only had 66 mhz processors and they 
were a HUGE bottleneck for RAID 5 performance.  I wouldn't be at all 
surprised if there is still a similar performance gap.

Pat