>But today, having 100 linux servers share the same read/only
>disk, and all 100 cache the same disk is undesirable when
>MDC could do it so well for all of them.

This is going to improve in the near future.

As for each Linux buffering its private data, I think that
is the question of global vs local optimisation. As long as
VM has plenty of storage I don't mind Linux cache its data.
But when VM runs out of storage and the Linux guest starts
to get page faults there is a lot of damage.
What I was thinking about is a device driver that talks to
CP and sees when the guest starts to get page faults. It can
then allocate memory in Linux (and not use it) to reduce the
footprint of the guest. I think that would force Linux to
give up I/O buffers sooner. When the processes in Linux need
more memory it would cause some swapping but it would swap
out those pages that I allocated (but don't reference) and
thus do little harm.
Instead of talking to CP it might be possible to watch the
pseudo page faults (but that might require a change in Linux
to also use it in kernel mode).

Rob

Reply via email to