Richard Troth wrote:
Barton might have some hard numbers, since Velocity has been pushing
XIP for quite a while now.
The benefit is academically obvious. To get your own measurement,
you'll need a multiplicity of guests and you'll want to measure VM
host load (esp paging) and Linux guest throughput (esp starting
execution of programs from XIP space). From the Linux side, execution
should be instantaneous. It may help to squeeze memory on both VM and
Linux (for the sake of your test), so you may want to run it in a
memory constrained test LPAR.
One thing I did do was modify the dcssblk module to count the number of
mapped page ops and the number of copy ops performed and exposed that in
/sys. That way I was able to see the activity change when mounted
with/without the "xip" option. Mainly to prove that I'd done things
correctly.
You gave me an idea though. I'll setup say 10 or 20 guests under a
second level VM. That way I can more easily see memory footprints
without normal work getting in the way. I realize it'll run like a dog,
but that's fine since we know that mapping memory will always be faster
than copy operations (even from cache).
Thanks,
Leland
