Process placement entails selecting a suitable node for a given process to
execute. In other words it is simply a remote execution facility. Migration
involves interrupting a running process and restoring it on a different
computer. Migration is rather more costly than placement, since the amount of
state and the complexity of obtaining it increases significantly once a process
begins execution. In addition, the whole process of migration involves a
significant amount of computational overhead, so determining whether
there is likely to be a performance improvement is non-trivial.
Krueger and Livny [#KruegerACOPA##1#] investigated whether the addition of a
migration facility to a distributed scheduler already capable of process
placement would significantly increase performance or not. They found that,
whilst placement alone is capable of a large improvement in performance, the
addition of migration achieves considerable additional improvement in many
cases. The magnitude of the performance improvement is dependent on several
conditions:
- There should be a high level of utilisation over periods long enough to
affect a user's perception of the performance of the system.
- Process initiation rates should be heterogeneous.
- The file system should store significant amounts locally, with little
replication at other nodes.
- The overhead of migrating a process should be high relative to its
service demand.
This conclusion is, however, contentious. According to Eager <#2759#> et al.<#2759#>
[#EagerTLPBO##1#]:
- There are probably no conditions under which migratory load balancing
could yield major performance improvements beyond those offered by process
placement. This is particularly true with respect to the advantages of systems
utilising process placement over systems which do not provide any load
balancing.
- Migratory load balancing can offer modest additional performance
improvements only under fairly extreme conditions. These conditions are
characterised by high variability in both job service demands and the workload
generation process.
- The benefits of migratory policies are not limited by their costs, but
rather by the inherent effectiveness of non-preemptive load balancing.