Re: Performance question, RAID5

["Keld Jørn Simonsen" <keld@keldix.com>]

On Tue, Feb 01, 2011 at 07:42:57AM +1100, NeilBrown wrote:
> On Mon, 31 Jan 2011 14:11:31 +0100 Keld Jørn Simonsen <keld@keldix.com> wrote:
> 
> > On Mon, Jan 31, 2011 at 09:37:46AM +0000, Mathias Burén wrote:
> > > On 31 January 2011 08:52, Keld Jørn Simonsen <keld@keldix.com> wrote:
> > > > If your intallation is CPU bound, and you are
> > > > using an Atom N270 processor or the like, well some ideas:
> > > >
> > > > The Atom CPU may have threading, so you could run 2 RAIDs
> > > > which then probably would run in each thread.
> > > > It would cost you 1 more disk if you run 2 RAID5's
> > > > so you get 8 TB payload out of your 12 GB total (6 drives of 2 TB each).
> > > >
> > > > Another way to get better performance could be to use less
> > > > CPU-intensitive RAID types. RAID5 is intensitive as it needs to
> > > > calculate XOR information all the time. Maybe a mirrored
> > > > raid type like RAID10,f2 would give you less CPU usage,
> > > > and the run 2 RAIDS to have it running in both hyperthreads.
> > > > Here you would then only get 6 TB payload of your 12 GB disks,
> > > > but then also probably a faster system.
> > > >
> > > > Best regards
> > > > keld
> > > >
> > > 
> > > Hi,
> > > 
> > > It's interesting what you say about the XOR calculations. I thought
> > > that it was only calculated on writes? The Atom (330) has HT, so Linux
> > > sees 4 logical CPUs.
> > 
> > Yes you are right, it only calculates XOR on writes with RAID5. 
> > But then I am puzzled what all these CPU cycles are used for.
> > Also many cycles are used on mirrored raid types. Why?
> > Maybe some is because of LVM? I have been puzzled for a long time why
> > ordinary RAID without LVM need to use so much CPU. Maybe a lot of data
> > sguffling between buffers? Neil?
> 
> What is your evidence that RAID1 uses lots of CPU?

Much of this is raid10, but it should be the same:
http://home.comcast.net/~jpiszcz/20080329-raid/
http://home.comcast.net/~jpiszcz/raid/20080528/raid-levels.html

It seems like cpu usage is rather proportionate to the IO done.
And the CPU usage does get up to about 40 % for reading, and
45 % for writing - this is most likely a significant delay factor.
For slower CPUs like the Atom CPU this may be an even more significant
delay factor.

For RAID5 the situation is even worse, as expected.

best regards
keld
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html