>> Well, there seems to be some disagreement here.

> And some misunderstanding as well. When Jim talks about multiple paths
> for the I/O, it is multiple paths to the control unit (that has a
> bunch of devices). So you can have multiple I/O's active at the same
> time to different devices on the same control unit. The operating
> system can get this without knowing much about it.

Right, I was answering a different question (what I thought was asked).

> The other issue is whether you can have multiple I/O's at the same
> time to the *same* device (when the read could be satisfied from CU
> cache while you wait on the write to complete). Provided the channel
> program clearly indicates its intentions you could, but I understand
> you need PAV. VM itself does not support PAV, but I think you can
> allow the guest to do so on dedicated DASD devices, right?

The lack of Parallel Access Volumes (PAV) in z/VM is a problem Barton
Robinson has pointed out here before. PAV allows multiple concurrent
I/Os to the same volume. With large volumes and a high I/O rate, the
lack of PAV support would be a performance issue. By using LVM to stripe
a large logical volume across several smaller "physical" volumes
(devices) you would get the best of both worlds as you will benefit from
the ability of the I/O subsystem and z/VM do perform I/O against
concurrent devices on the same control unit. Of course, since the real
disk subsystem (say in IBM Shark/ESS or a StorageTek SVA) is already a
RAID 5 device with striping, I don't know if the end result of LVM
striping and multiple logical paths would give much benefit (as someone
pointed out earlier today).

I hope I got this right now.

Regards, Jim

Reply via email to