Load balancing has been shown to have the potential to provide better
performance than load sharing if the overhead of load distribution is ignored
[#KruegerTDOOD##1#]. It is rather more equitable than load sharing and it
addresses situations in which some processors are lightly loaded and some
heavily loaded; it does not restrict migration to idle processors.
According to Goscinski [#GoscinskiDOSTL##1#] there are a number of questions
which must be answered in forming any load balancing policy:
- When to migrate processes to balance the system workload?
- How to compare workloads of different computers to decide which should
be offloaded first whilst others wait?
- Which process from a chosen computer should be moved?
- Which computer is the best destination for a process to be migrated?
- Which computer should be involved in searching for a lightly loaded (or
idle) computer? Should it be the source computer for a given process, or should
lightly loaded computers perhaps offer their services?
- When are the scheduling/load balancing decisions to be made?
- Which parameters should be taken into consideration in making the
decisions required by the above mentioned questions?
- What can happen when data is not available or out of date?
- Should the data necessary to make these decisions be gathered and
stored centrally or in a distributed manner?
- Should computers cooperate in making decisions or not? If so, what
mechanisms are needed to update system state in relevant computers?
- What is the tradeoff between performance and overhead of making
scheduling/load balancing decisions?
- How can one avoid overloading a lightly loaded computer?
- How can one avoid starvation by constantly migrating the same processes
between machines?
There are several different reasonable answers to all of these questions.
Consequently there are tens if not hundreds of different approaches presented
in the literature. This means that it would be impracticable to cover each in
any detail. Fortunately, they can be separated into a relatively small number
of different categories, which can be discussed separately. The taxonomy in
general use is that due to Casavant and Kuhl [#CasavantATOSI##1#].