They finished, Microwiper counts the dirtied pages

They
propose Microwiper that can efficiently propagate memory in live migration of
virtual machines. Microwiper is based on two strategies: ordered memory propagation
and transfer throttle. Within VM memory, page rewriting rates differ greatly.
For pages with higher rewriting rates, the chances that they are to be written
are very high. Sending those pages will be likely to make them dirtied again
and then retransferred. On the contrary, the pages with lower rewriting rates
are good candidates to transfer in priority. Instead of counting rewriting rate
of every page, they group pages together. Memory stripe is a multiple of
consecutive VM memory pages. VM memory is thus divided into many stripes of
equal length. Reorganizing VM memory into a series of stripes has the benefit
that rewriting rate of a stripe is more stable (this can partly be explained by
the principle of spatial locality) and much less overhead than that of a single
page. Stripe size is decided by the size of allocated VM memory size, e.g., if
VM memory size is 128MB, then stripe size could be 32 pages, etc. They have a
simple method to statistically sample the rewriting rates of memory stripes.
Every time an iteration is finished, Microwiper counts the dirtied pages for
each stripe in that iteration. This number is then divided by the time the
iteration lasts. The result is the rewriting rate of the stripe. Finally, they
apply ordered propagation at the granularity of memory stripe: in each
iteration, they calculate rewriting rates of all memory stripes, sort them in
ascending order, and then dirty pages in stripes are transferred successively
according to that order.

Since
VM memory can be updated much faster than network transfer rate, it should be
avoided to send pages whose chances for not being updated while being
transferred are low. Writable Working Set (WWS for short), which
is a set of pages that are written often and should be transferred in final
stop-and-copy phase.WWS is the lower bound to send in migration downtime,
provided it must be accurately identified. However, current migration
approaches have not properly dealt with this issue. We take into account two
factors in forming it. One is rewriting rate for pages in WWS must have the
highest rewriting rates. The other is available network bandwidth between
source and destination hosts, which is dynamically estimated and serves as a
threshold (available network bandwidth is the actual bandwidth that can be used
by sender). We thus derive another strategy transfer throttle to
optimize memory propagation: dirty pages are transferred until their added rewriting
rates exceed estimated available bandwidth; transferring of the left dirty
pages is throttled and the next iteration is started afterwards.

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!


order now

 

In each iteration, they first calculate the
rewriting rates of memory stripes from the previous iteration, then the memory
stripes are sorted by their rewriting rates; They transfer dirty pages in
stripes according to that order and estimate network bandwidth dynamically;
rewriting rates of transferred stripes are added up and compared with estimated
network bandwidth; at the time the accumulated rewriting rate exceeds the
estimated bandwidth, the next iteration is started immediately.6