From modes to patterns: Pattern-based resource management in time-critical applications

Abstract

Resource management is a vital activity of many resource platforms. For time-critical applications the principle resource to be managed is processor time. For many streaming video applications processor usage of their individual components follows a limited set of modes each of which represents a small range of processor utilization values. In this paper, we show that these modes often follow specific patterns which can be detected by monitoring processor usage at runtime. Furthermore, a cost-effective pattern detection algorithm is presented and a class of strategies is defined that use patterns to predict future resource usage. These strategies are capable of extending the reservation period beyond the next mode, which is the standard for mode-based resource management, Thus, not only management effort is reduced, but also the quality of the reservations is increased. To determine reservation quality, metrics are used that measure the extent of both over- and under-provisioning. The applicability of the detection method and strategies is illustrated through a set of experiments. One set of experiments shows the existence, rapid emergence and ease of detection of patterns. Another set of experiments demonstrates the reservation quality for several strategies, and indicates the dependence of that quality on the parameters used to select the strategy from the defined class. Thus it is shown that pattern-based management provides a cost-effective and accurate means to manage processor utilization of individual components and therefore can be used both for intra-and inter-application resource management.