Apologies for the new questions - please enlighten
me.
In the traditional Linux kernels, we have the delayed I/O
concept to improvement performance. The disk block I/O request is
bunched together whenever possible with the previous block request, for
performance reason.
Analogously, due to the large overheads in making hypercalls,
is it possible that we can do this? Ie, the core instructions will
still be executed in its original order, but the overheads of making the
multiple VM exit/entry are all bunched together and executed once.
The hypercalls necessarily are coming from different CPU, right? Can
further improvement be made by removing the atomicity requirements of the CPU
instruction that triggered off the VM exit condition? (sometimes at
least?) If so then it may be possible to bunch together hypercalls
from the same CPU as well.
Similarly, for the Xenbus state transition machine - can we
improve performance in inter-domain communication - but not
necessarily satisfying all Xenbus request immediately all the time?
Can it be done both for PV and HVM scenario?
|