[clug] forget RAID?

[Anthony David]

On Wed, Feb 21, 2007 at 08:23:45PM +1100, Michael Cohen wrote:
> Dale,
>    The parity calculations are actually not significant on modern CPUs - most
>    software raid implementations are actually much faster than most hardware
>    implementations (which offload the parity to the hardware). 

True.
> 
>    Raid 1 is a mirrored configuration which means that on reading consecutive
>    blocks - you can actually read one block from each disk before waiting for
>    the read to complete - hence raid 1 gives 2x increase on read speed.
> 
>    Writing on the other hand is still 1x because you still need to write both
>    blocks to both speed - hence why raid 1 is great for reading - not so good
>    for writing.

Surely your OS write can wait for one write to disk to complete. The other
one can finish when it likes.

> 
>    Raid 5 OTOH is less good for reading because for a 3 disk investment you
>    only get 2x speed reads (only 2 sectors can be done simultaneously). But
>    writes are not quite so bad because you can start writing 2 disks at once
>    and then write the parity a little later (so 1.5x speed increase).

RAID 5 is better for reading, especially as your stripe size goes
to 5, 6 whatever you are comfortable with and fits your application
profile. Writes are a real burden, especially with software RAID 5
as you are "flying without a parachute". Run a RAID 5 controller with
no write cache and see how fast your writes go. It has to do two
writes that HAVE to complete. Software RAID 5 has the same problem.

RAID 1+0 gives you the read performance of
RAID 0 and the reliability and write performance of RAID 1.

You do need a lot of disks though. Personally, I always felt
the "I" in RAID was only applicable if someone else was
paying for the disks. 

Sometimes you need disk I/O performance above that available from a
single drive.

Many times you need filesystem sizes exceeding that of a single drive.

You are going to have to stripe your drives to get that sort of
performance and/or capacity. You then increase the risk of data
loss. RAID is a technology to address this risk and certainly has
its place.

PS I do feel a little sheepish at only having read the first page
or so of the google paper. I have been busy fixing hardware,
including disks :-) Damn. First week of lectures and have missed
half of them already :(

> 
>    Michael.
>   
> On Wed, Feb 21, 2007 at 05:23:07PM +1100, Dale Shaw wrote:
> > Hi Michael,
> > 
> > Just a minor point:
> > 
> > On 2/21/07, Michael Cohen <michael.cohen at netspeed.com.au> wrote:
> > >For reading raid 1 is great, not so good for writing.
> > 
> > I think you'll find it's RAID-5 that suffers from heavy write
> > activity. There are no parity calculations for RAID-1.
> > 
> > cheers,
> > Dale
> > 
> -- 
> linux mailing list
> linux at lists.samba.org
> https://lists.samba.org/mailman/listinfo/linux

-- 
Anthony David

Gambling(n): A discretionary tax on those asleep during high school maths
http://adavid.com.au/
0xA72CE1ED fingerprint = EA1E C69E FE59 BBE1 AA4B  F354 BD09 9765 A72C E1ED