Condition Evaluation for Speculative Systems: a Streaming Time Series Case

Application systems often need to react with certain actions whenever some preset conditions are satisfied. In many cases, the evaluation of these conditions takes long time, but some prediction of the results can be obtained rather quickly. In this situation, speculation may be a good idea. That is, the system takes predictions (speculation) to prepare (such as prefetch) for the possible reaction. Obviously, the risk is the waste due to false alarms. Higher precision prediction results in less waste, but takes longer time and may reduce/eliminate the opportunity for speculation. A balance needs to be struck. A qualitydriven prediction subsystem is thus necessary, so that the “user” of the prediction subsystem can impose quality (in terms of precision and response-time) requirements. This paper focuses on such a prediciton subsystem with conditions on streaming time series. Two problems need to be solved: how to predict the precision and how to achieve the required precision in an optimized way. The paper introduces a prediction model to tackle the first problem, and presents an algorithm to attack the second. Experiments show that the prediction subsystem works well.