[CentOS] Re: Read performance on software RAID1?

Thu Apr 12 19:57:03 UTC 2007
WipeOut <wipe_out at users.sourceforge.net>

John R Pierce wrote:
> WipeOut wrote:
>>> AFAIR asks for the data from the raid device, and the md code sends the
>>> request to both drives. The drive that replies with data first is the 
>>> winner.
>>> So the system would be reading staggered by that method. Now writes 
>>> are a
>>> different story.
>>
>> Ok, so in theory then read performance on software RAID1 may be better 
>> than a single disk but how much better would depend on the hit rate of 
>> the requests being distributed..
>>
>> Whats the story on writes?
>> You make it sounds like there is something happening there that you 
>> are not happy about..
> 
> writes have to be issued to both drives.  assuming the drives are on 
> seperate channels, or on a shared channel thats fast enough not to block 
> on two current accesses, then writes should be no slower than a single 
> drive.
> 
> his description of reads is somewhat imprecise.    the MD code will send 
> the request to the drive thats least busy, not to 'both drives'.     I 
> believe it also does some elevator seek optimization (eg: if both drives 
> are idle, it will go to the one that was last accessed closest to the 
> new request).

Ok.. That all sounds logical.. Thanks for the feedback.. I figured it 
would do something to make use of the two copies of the data but just 
wanted to check.. :)