smart raid1 reads?

smart raid1 reads?

[Nicholas Leippe]

Hi,

I searched the archives, and the last time this was mentioned was in 2001, and 
no response was ever given.

In a raid0, of course, reads pull off every disk in parallel, thus performance 
scales with # of disks.

However, with raid1, it does not do this--it always reads from just one half 
of the mirror.  Why doesn't it read from both halves of the mirror in 
parallel?  I would expect it to, and would expect raid1 to perform similar to 
raid0 for reads.

I have observed this with 2.4.18 and 2.4.20 thru at least 21rc2.


Nick
please cc

Re: smart raid1 reads?

[Michael T. Babcock]

>
>
>In a raid0, of course, reads pull off every disk in parallel, thus performance 
>scales with # of disks.
>
>However, with raid1, it does not do this--it always reads from just one half 
>of the mirror.  Why doesn't it read from both halves of the mirror in 
>parallel?  I would expect it to, and would expect raid1 to perform similar to 
>raid0 for reads.
>  
>

I was looking for more information on why this is myself today.  Its 
quite possible that the complexity is too great for anyone to be willing 
to try it yet.  After all, if you want to read off of disk 0 and disk 1 
at the same time, you need to be sure the data you're reading doesn't 
include any blocks waiting to be sync'd from the other drive.

I don't know what the in-memory structures look like for 'blocks waiting 
to be sync'd', but depending on their design, this could be easier said 
than done.

-- 
Michael T. Babcock
C.T.O., FibreSpeed Ltd.
http://www.fibrespeed.net/~mbabcock

Re: smart raid1 reads?

[Michael T. Babcock]

Nicholas Leippe wrote:

>I don't know about the internals either, but it's already issuing block 
>writes to both drives in parallel, I don't see why synchronization is an 
>issue.  Block layers don't usually care about serialization--that is left  to the filesystem and higher os levels, no?  If so, I would think it shouldn't matter--md1 is in effect here the new block layer--merging two other block devices into one--nothing should be bypassing that layer for any blocks in the partition it's using.
>  
>

Part of what bothers me is that even without parallelizing for one large 
read, if there are multiple queued reads from one RAID1 device that 
aren't nearby (see elevator code), then read the second set from the 
second drive, the third set from the first drive, and so on to minimize 
head movement.

-- 
Michael T. Babcock
C.T.O., FibreSpeed Ltd.
http://www.fibrespeed.net/~mbabcock

Any mdadm based cgi perl script for webmin?

[Bo Moon]

Hello,

I downloaded the latest WEBMIN, but its RAID tools are based
on the RAIDTAB config and old tool.

I want to use new MDADM tool, but I am not an expert on PERL.

If anyone already modified it or know where some references are,
May I get some new cgi perl program(or prototype) which use MDADM for
webmin?

Those are under /usr/share/webmin/raid:
               create_raid.cgi,  raid_form.cgi, raid-lib.pl, ......

Thanks in advance,

Bo