Sorry for the delay:
The viability of virtual machines (and migrating them) for distributed systems
applications
- Migrating (at least blindly) is probably not a good idea for systems that
are supposed to be fault tolerant, because if your n replicas of your
data/service are on fewer than n physical machines, a small number of node
failures could take down the entire system.
- The migration itself is fairly feasible with respect to performance when
done inside a data center, as computation does not need to freeze while
the gigabytes worth of VM state is copied over the network.
- Migration is probably not feasible on an Internet scale.
Other approaches to power saving
- Fix the hardware
- Build your data centers in Alaska
The optimization problem
- Is a simple bin packing problem (minimize the number of physical nodes while
meeting capacity constraints on the amount of CPU power and memory a
physical node has) enough to express the relevant complexities, or would a
more descriptive (and thus more difficult to solve) formulation provide a
noticeable difference?