The previous section argues that the management of all primary resources is important to the performance of any distributed system, therefore I suggest the following hypothesis:
This process mix approach to load distribution is a logical extension of the traditional load distribution philosophy noted in Casavant and Kuhl [CK88], `` being fair to the hardware resources of the system is good for the users of that system''.
This hypothesis has two major implications:
A priori knowledge about a job for initial placement can be provided by estimating its future behaviour from its historical behaviour. Examples of prediction systems providing a priori knowledge for initial placement are Bond [Bon93] with a heterogeneous system, and Goswami et al. [GDI93] in a homogeneous system.
No prediction mechanism can expect 100% accuracy with a dynamic workload, and therefore the question arises of how sensitive the performance of a policy is, when relying on such predicted values.
As prediction can provide a priori knowledge to load distribution policies, it should be possible to place a job correctly, and therefore it will be unnecessary to later migrate that job after it has begun execution. A job that is correctly placed is one that behaves as predicted and results in resource demands that are consistent with the policy by which it was placed.
Workloads containing interactive processes suffer a degree of randomness and therefore predictions are not always accurate. In particular, any system that relies totally on prediction is likely to suffer from this degree of randomness that can not be accounted for by the load distribution policies. Therefore hosts should still be monitored to ensure they are consistent with their computed loads.
If the initial placement policies using prediction are indeed sensitive to this inaccuracy, hypothesis 3 may lack support. One way around this, is to accept that mistakes are made with initial placement, and use another mechanism such as migration to provide a degree of correction. Many systems, such as Sprite [DO91] and V [TH91] provide both initial placement and process migration systems, yet do not integrate the use of both mechanisms in a single load distribution policy.