Re: RAID 6 mdadm

[Stan Hoeppner <stan@hardwarefreak.com>]

On 4/15/2013 1:44 PM, Dexter Filmore wrote:
> Am Wednesday 10 April 2013 23:54:28 schrieb Stan Hoeppner:
>> On 4/10/2013 12:32 PM, Dexter Filmore wrote:
>>> Am Wednesday 10 April 2013 18:05:08 schrieb Ross Boylan:
>>>>> You cannot get more than about 200MB/s out of a RAID1 setup, not even
>>>>> with
>>>>> 15krpm SAS drives. the RAID1 will never be faster than a single disk.
>>>>> How
>>>>> would it.
>>>>
>>>> It can read faster than a single disk by combining reads from different
>>>> sectors on different disks.  I think your assertion is incorrect for
>>>> reading, though  not for writing.
>>>>
>>>> Ross Boylan
>>>
>>> Then dm would have to manage/combine the reads from the involved disks.
>>> I do not know better, but I doubt this is done.
>>> I'll ask the raid list for curiousity's sake.
>>
>> It's not necessary to ask on linux-raid, and as you have not yet done
>> so, I'll clarify this now:
>>
>> Both the md RAID1 and RAID10 personalities will read sectors in parallel
>> from both disks in a mirror pair in most circumstances.  This has been
>> the case for many years.  This is a read optimization only.  Writes
>> occur in parallel as well, obviously, but as it's the same data the net
>> data throughput isn't doubled, only the raw device throughput.  This
>> should be obvious to anyone.  I'm merely being complete in my description.
>>
>> --
>> Stan
> 
> Interesting. Only 1 and 10 or 5/6 as well?

Hardware, i.e. firmware, and many software implementations of RAID1 and
RAID10 typically only read from one drive in the mirror pair and write
duplicate blocks to each.  This means that one drive is always idle
during read operations, but both drives are active for writes.

md/RAID takes advantage of the idle drive in a mirror pair by reading
ahead in the queue and then splitting the reads among the two drives,
*in addition to* performing striped readout.  All other things being
equal, md/RAID1/10 reads much faster than other RAID1/10 solutions.  As
always the magnitude of the increase depends on the workload and access
patterns.

RAID5/6 don't have mirrored drive pairs, thus don't have any idle disks
during reads, as RAID1/10 normally do.  So no, this doesn't apply to
RAID5/6.

-- 
Stan