[dm-devel] About using multipath in SLES 10

[malahal at us.ibm.com]

Erez Zilber [erezz at voltaire.com] wrote:
> I understand. I've tried disktest, and it work ok. I have another
> question - I'm running a single iSCSI initiator against 2 iSCSI
> target. I started disktest on /dev/dm-1 and after a few seconds, I
> disconnected the primary target. It took ~2 minutes until it failed
> over to the secondary device. During these ~2 minutes, it seemed that
> disktest is still reading data from the target, just slower. That's
> very strange. After failover was completed, I was able to use the
> secondary device.

You said, "it seemed that disktest is still reading data from the
target". Maybe it is reading some cached data (read-ahead or some other
such thing...) Also, the multipath kernel module would not know that you
have disconnected the target until an I/O fails. Depending on your error
injection and the susystems design, the I/O failure could be as a result
of a timeout (this depends on your/distro setting, generally 1 minute).

> I have 2 questions:
> 
> 1.	> 	Which parameter in  multipath.conf do I need to change
> in order to failover in a few seconds? Is it polling_interval? I saw
> that the default value is 5 seconds, which should be ok.

Not an expert, but polling interval can't change the failover time. It
may change the 'failback' time though!

> 2.	Before multipath decided that it needs to failover, why did I
> see that traffic is still running? It had no device to work with at
> that time.

Depends on 'how you saw the traffic'! Could be false alarm.