Re: install Etch with raid level 10
Goswin von Brederlow wrote:
You should do some speed tests first. raid1+0 (or lvm stripes) behave
differently to raid10 in its different layouts.
Having just finished a bunch of raid10 speed testing I couldn't agree
more. Test it before making assumptions. Many factors go into what
overall speed you'll get from a raid10. These are the disks themselves,
the controller(s), number of disks, chunk size, raid layout (near,
offset, far), the io elevator (scheduler), and ultimately what you use
the system for. I've seen a 200GB disk where write speed (single) is
slightly faster and a 300GB disk from the same manufacturer and family
that seemed to prefer read over write. These small differences were
reflected in the overall raid speed. With controllers I had very poor
performance from a certain "server" SATA card, and much better with a
noname SATA card (140-250% of the other, depending on what you're
doing). For me, write speed (and sequential) was important. The near
layout seems to provide the fastest write speed at some expense of read
speed. The far layout has worse write speed, better read speed, but can
vary widely depending on the io elevator. The offset layout was a
compromise with writes, and the slowest with a single read. However,
with two or more simultaneous reads it was by far the fastest (single at
139MB/s, two at 130MB/s and 122MB/s), so you should consider (and test)
it if you ever read more than one thing at a time from your disks... :)
After testing and seeing pitiful write performance, I'm dropping raid5
and going with a six disk raid10, near 2 layout, 256k chunk size with
the deadline elevator. My application is disk-to-disk backup.
In short, test. But this can take a lot of time. If you can't afford
extensive testing time, try something, test it, then decide if it's
acceptable. Also, at least try the different io elevators (with your
real workload) as this is a runtime change and doesn't require any
change to raid config.
--
-Cedar
Reply to: