SHF:Small: Accurate and Computationally Efficient Predictors of Java Memory Resource Consumption

点击翻译按钮获取中文摘要

The Java programming language is widely-used and of great commercial and economic significance. It is favored in part because it features automatic management of the computer memory resources it uses, simplifying such management for the programmer. Memory management in Java (and other managed languages) has reached a plateau in cost and effectiveness because most current techniques are tuned based on a small number of coarse-grained measures gathered while programs run. Substantial improvement might be gained from using more accurate estimation of current and near-future memory use to drive better memory management decisions. This would reduce the time, memory, and energy requirements to run Java programs. This is of significance to the full range of Java applications from small embedded systems through laptops and desktops to large servers. There is therefore an urgent need for techniques to derive better online predictors of Java memory use.The long-term goal of the research program this award will support is to substantially improve memory allocation and garbage collection effectiveness by using better online predictors to drive more sophisticated allocator and collector decisions. The objective of this particular project is to develop machine learning techniques that induce accurate and computationally efficient predictors of characteristics of Java memory allocation that influence memory manager performance. Examples include predicting the volume of objects that become "garbage" (can be reclaimed and reused for future allocations), as well as objects that will be in use for a long time and will not become garbage soon. The approach is to learn models that predict memory usage based on features compiled from observable run-time events like calls to particular methods or allocations of certain objects. Data to learn models will be obtained from analysis of detailed program execution traces. Features will be selected that are both informative of memory use and computable with low space and time overheads. Programs will then be modified to compute these features as they run, and real-time predictive models will be used to predict future memory usage as programs execute. These predictions will be used to improve memory management performance. This will be accomplished by, for example, improving the timing of garbage collection so that it occurs at points during program execution that result in higher memory reclamation with lower effort.