zfs discuss - Sep 2006 - please remove my ignorance of raiding and mirroring

Folks,

I''ve heard that for small reads/writes (like a mailstore), mirroring  
is preferred to raiding.  Can some explain why or direct me to that  
info?  I assumed that raidzn would be the preferred method but  
apparently not.

thanks,

keith

On September 29, 2006 11:09:21 AM -0500 Keith Clay <clayk at acu.edu>
wrote:
> Folks,
>
> I''ve heard that for small reads/writes (like a mailstore),
mirroring  is
> preferred to raiding.  Can some explain why or direct me to that  info?
> I assumed that raidzn would be the preferred method but  apparently not.
<http://blogs.sun.com/roch/entry/when_to_and_not_to>

Mirroring is more efficient for small reads (the size of one ZFS block or less)
because only one disk has to be accessed. Since RAID-Z spreads a ZFS block
across multiple disks, and the data from all disks is required to verify the
checksum, every read accesses every disk.

Mirror read:
  1. Pick one of two (or more) disks.
  2. Read from that disk.
  3. Verify that the data is correct.

RAID-Z read:
  1. Determine which disks hold the requested data.
  2. Read from all of them.
  3. Verify that the data is correct.

Small writes, or larger writes which do not exactly line up with a ZFS block,
require a pre-read of the old data. As for other reads, this requires accessing
all disks in the RAID-Z case. Large writes, however, should be more efficient
using RAID-Z instead of mirroring, as less total data is written to disk
(data+parity, instead of data+data), and a pre-read is only required at the
beginning and end of the range being written.

This is a difference from RAID-5. In RAID-5, small reads are more efficient than
mirroring because there are more disk spindles available, and a small read
usually uses only one. Small writes are less efficient than mirroring in RAID-5
because they require a pre-read phase (though only to two disks, rather than
all).
 
 
This message posted from opensolaris.org

Anton B. Rang wrote:
> This is a difference from RAID-5. In RAID-5, small reads are more efficient
than
> mirroring because there are more disk spindles available, and a small read 
> usually uses only one. Small writes are less efficient than mirroring in
RAID-5
> because they require a pre-read phase (though only to two disks, rather
than all).
I don''t believe this statement is generally true.  Consider the case of
4 disks.
RAID-1+0 could service 4 independent read iops concurrently.  RAID-5 could only
service 2 independent read iops concurrently.  More generally, the 2-way mirror,
RAID-1+0 could always service at least 2 (>= N/2) independent read iops
concurrently while RAID-5 could only service a maximum of 2 (<= int(N/2))
concurrent read iops.
  -- richard

I was perhaps slightly hasty; RAID-5 is more efficient than a 2-disk mirror for
reads, but RAID-10 (striping & mirroring) can match (or slightly exceed) its
efficiency for small reads. In both cases, there is a single disk accessed.

With 4 disks, RAID-1+0 can service four independent reads concurrently, while
RAID-5 can service up to four independent reads concurrently, but perhaps as few
as 1 if all reads are directed to the same spindle. Under heavy loads, all
spindles will generally be busy, so performance is similar.
 
 
This message posted from opensolaris.org

On Oct 1, 2006, at 11:24 PM, Anton B. Rang wrote:
> I was perhaps slightly hasty; RAID-5 is more efficient than a 2- 
> disk mirror for reads, but RAID-10 (striping & mirroring) can match  
> (or slightly exceed) its efficiency for small reads. In both cases,  
> there is a single disk accessed.
>
> With 4 disks, RAID-1+0 can service four independent reads  
> concurrently, while RAID-5 can service up to four independent reads  
> concurrently, but perhaps as few as 1 if all reads are directed to  
> the same spindle. Under heavy loads, all spindles will generally be  
> busy, so performance is similar.
>
Let me see if I understand.  If I have a RAID10 with 4 or more disk,  
then for a mailstore this would be the optimal solution.  I expect to  
have 6 or more disk in mirrors, so would that increase it''s efficiency?

keith