"The Rate Monotonic Scheduling Algorithm: Exact Characterization And Average
Case Behavior" by Lehoczky, Sha, and Ding attempted to mathematically
characterize the rate monotonic algorithm, shown by Liu and Layland in 1973 to
be the most optimum static priority scheduling algorithm, as well as to
present a Monte-Carlo-style statistic model of its behavior. Lehoczky, Sha,
and Ding showed that under a certain set of simplifying assumptions (listed in
the introduction of Section 2, Problem Formulation), the rate monotonic
algorithm can successfully schedule all tasks if the maximum of the tasks'
cumulative processor demand index (some math notation and definitions I can't
type here) is at most 1. The authors then go on to test the algorithm's
average behavior using stochastic methods: tasks with random periods and
random processor utilizations are generated and then scheduled with rate
monotonic; the utilizations are then scaled up until the algorithm can no
longer successfully schedule all the tasks. The tests showed that as the
number of tasks was increased, the maximum or "breakdown" utilization level
approached a constant that depended on only the probabilistic distribution of
the tasks' periods and not the distribution of the tasks' processor
utilizations.
The authors' paper, published in 1989, should be applauded for finally
addressing the exact mathematical characterization of the rate monotonic
algorithm that Liu and Layland had shown in 1973 to be optimum. Though the
paper made several simplifications and assumptions in both the mathematical
characterization and the stochastic modeling, the authors clearly stated what
assumptions had been made and how they might be better addressed in the
future. In Section 5, Summary and Conclusion, the authors also listed many
ways in which the results of the paper should be generalized for application
in less restrictive circumstances.
This was an interesting paper that allowed me to understand rate monotonic in
more detail (though I admit I knew nothing about it to begin with). I was in
a little over my head with the mathematical details, but the general results
the authors presented were very neat and surprising. In particular, knowing
nothing about rate monotonic beforehand, I would have expected intuitively
that its asymptotic behavior depended on the task set's utilization
distribution (not its period distribution), rather than the other way around.