LVM - Mirroring write impact

[HMC]

Doing some array testing using an iozone command, iozone -s 1G -r 32K -t 32 -i 0 Write only - Unix. Using LVM, results as follows: logical volume 1 local half mirror - 157mB/s
1 remote half mirror - 95MB/s
Fully mirrored - 52MB/s
Why the big drop ?
Another result with both sides of mirror local
local tests - 101MB/s and 111MB/s
mirrored test - dropped to 12MB/s (really bad!)

Is this what we should expect ?

Similar tests on a HP array with vxvm did not give the same results - is there a reason for this ?

[Torsten.]

What exactly means "remote"?

[Tim Nelson]

The writes do have to go to both disks. Are both disk on same controller ?

As far as the write to an array, most arrays write to cache hence the faster times. The array could also be raid0+1 vs raid1(not mentioned here so it is not an apples to apples comparison).

[HMC]

Two different HBA's out to two switch ports, which for one test was to two ports on a local array. For the other test one port was on the local array the other was across the SAN (via dwdm - 20KM) to another array. Although as the results showed the remote site element doesn't seem to make a difference.

[Torsten.]

My very first idea would be about a bad connection to the remote site. Did you already inspect the switch logs?

[HMC]

Can't be the remote link as we have the problem with the two luns local as well.

The question really is: If you mirror using LVM and unmirrored volumes give you 100-150MB/s throughput what would expect the throughput for a mirrored volume to be ? (each side of the mirror is on a different hba and different switch port and array port etc.).

[Tim Nelson]

I see the dillema now. Sorry I saw local and was thinking locally attached DASD disk.

Just from the stats below there certainly is a degradation in performance to your remote array ( it would make sense as any remote connection will always be slower ).

As far as guessing the added delay of almost 2x using LVM mirroring to the remote, yep it is a dillema.

Keep an eye on the response time stats when testing the remote scenarios, iozone could just be driving the ios over what the remote can handle.

Maybe array replication like EMC's RDF or whatever equivelent is available for your array type is a better solution than LVM mirroring ?