[linux-lvm] multiple volumes..

[Adrian Head]

On Wed, 10 Apr 2002 08:01, you wrote:
but that REQUIRES, that the disks are not
> hd<a,b,c,d,e,f,g,h>
> You should try to avoid having more than one (active) disk at the same
> controller at once.

I've always wondered about this statement.  

The idea AFAIK is that unlike SCSI, IDE has the master drive control the 
slave drive on the same controller.  
1) This causes problems where the drives are a different 
model/make/manufactuer as the master drive will downgrade the settings of the 
drives (master & slave) to the lowest common denominator; and therefore, the 
speed in this case is affected.  
2) Having the master drive control the slave drive also causes problems if 
the master fails as then both drives fail or if a drive takes down the chain 
both drives fail.
3) maybe a couple of others that I have forgotten at the moment.

When I was setting up my server at home 1.2GAthlon AsusA7V-133 384M RAM 9x 
IBM 40G drives, 1 Promise Ultra100 & 2 Promise Ultra100TX2 I did a couple of 
tests.  
I found by using a flaky drive from a previous life (a western digital that 
when it would die would take down the controller) that it didn't matter if 
the drive was a single master on a controller, a master on a shared 
controller or a slave on a shared controller.  When the drive would fail - it 
would take the whole machine down with it.  The Kernel would not die but it 
would deadlock waiting for interrupts to return and the only way to fix the 
issue was to hard reset.  Having the flaky dirve as a standalone drive or as 
part of a software RAID made no difference.  So I concluded that at least in 
my setup I gained nothing by only having one IDE drive per controller.

I also ran bonnie++ tests on a software RAID0 using 4 master only drives 
(hde, hdg, hdi, hdk) and 4 master-slave drives (hde, hdf, hdg, hdh).  The 
results of the test were that seek times for the master-slave case were half 
that for the master-master case.  Read times did drop but only by about 
.75M/s from 80M/s but the write times improved by about .25-.75M/s from 
50M/s.  So I concluded that at least for my setup it was better to have the 
master-slave case because the loss of read speed was made up by the increase 
in write speed.  

So my final configuration was that hda, hdb were system disks.  The RAID5 
used hde, hdg, hdi, hdk with a hot spare hdo and the RAID0 used hdm & hdn.

Now the tests weren't what I would call conclusive or proper scientific tests 
but I believe they were valid.  I've always wondered what others have found 
as what I found seems to fly in the face of common rules of thumb.

There is a specific mailing list "linux ide" 
<linux-ide-arrays at lists.math.uh.edu> that deal with large IDE arrays so you 
might like to also ask your question there and see what is suggested.

-- 
Adrian Head

(Public Key available on request.)