Dennis Wicks wrote:

Is there any thing in either VM or linux that would make the use
of multiple V-DISK swap spaces better or worse than a single one?

I don't mean just one versus two but more on the order of
ten 100 meg. swap areas versus one 1 gig. swap area. Or even
something like ten 50 meg. areas vs. one 500 meg. area.

Yes, there is. Look out for a Redbook that shows you real soon :-) Writing the full story in mail would spoil the day of most on the list, of my residency leader, and of myself. So just a few points.

(I am not sure this justifies a very large number of swap disks,
but it does show you why splitting in two or three makes sense).

1. When using the proper configuration (i.e. the Diagnose driver)
we can run a swap disk faster than you blink your eyes. Think of a 32 MB
Linux swapping at 48 MB/s and still give interactive response (this is
one order of magnitude faster than with the ECKD driver on 3390).
As Mark points out, on other platforms multiple swap disks are used with
*equal* priority to increase the bandwidth. Given the reported swapping
rates, I have neither confidence nor desire to go there for VDISK.

2. VDISKs allocated for your virtual machine are almost for free as long
as you don't touch the blocks (the blocks don't even exist on page space).
Even when you make an empty filesystem or swap disk on the VDISK you hit
only a few blocks and the cost is minimal. The difference between 2 empty
disks of 100 MB or one empty disk of 200 MB is minimal as well.

3. Just like Linux will ultimately use all memory that you give it, it
will also use all swap space that you give. Allocation on swap disk is
done in such a way that the active area (the hot spot) sweeps over the
disk to avoid (virtual) head movement, increase chances of block paging,
and just because you can. When CP cannot keep the entire VDISK in memory
it will page out the least recently used blocks, and bring them back in
when Linux replaces the contents. So in the end each page swapped out by
Linux will cause a page-out and page-in by CP. And it fills your page
packs.

So if you're serious about this, we would recommend multiple VDISKs with
*different* priority. That makes Linux use free blocks in the first one
before using the next one. Sizes depend on your workload and the usage
pattern of your swap disk. You might for example double the size of each
level of swap disk, and have the first as large as your virtual machine.
Your favorite peformance measurement tool can report how much of the
VDISK is being referenced, and where it resides. That should assist you
in reviewing the results of your planning.

Rob

Reply via email to