Hi,
Since the data is available in two places on a RAID1 array does the code in any way make use of that for improved performance by reading (different) data form both disks at the same time? Or is the one disk simply a slave that just receives copies of the data and is never really read from at all?
Thanks
WipeOut wrote:
Hi,
Since the data is available in two places on a RAID1 array does the code in any way make use of that for improved performance by reading (different) data form both disks at the same time? Or is the one disk simply a slave that just receives copies of the data and is never really read from at all?
reads are distributed to both drives. the kernel raid implementation is considerably more sophisticated about this than the typical bios/driver based fake raid.
WipeOut spake the following on 4/12/2007 9:38 AM:
Hi,
Since the data is available in two places on a RAID1 array does the code in any way make use of that for improved performance by reading (different) data form both disks at the same time? Or is the one disk simply a slave that just receives copies of the data and is never really read from at all?
Thanks
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.
Scott Silva wrote:
WipeOut spake the following on 4/12/2007 9:38 AM:
Hi,
Since the data is available in two places on a RAID1 array does the code in any way make use of that for improved performance by reading (different) data form both disks at the same time? Or is the one disk simply a slave that just receives copies of the data and is never really read from at all?
Thanks
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..
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).
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.. :)
-
John R Pierce -
Scott Silva -
WipeOut