Efficient Dynamic Memory Allocation in Data Stream Processing Programs

Data Stream Processing is a paradigm enabling the real-time processing of live data streams coming from sources like sensors, financial tickers and social media. The history of the stream is often maintained in sliding windows and analyzed to produce timely notifications to the users. A challenging issue in the development of parallel implementations of such computations is efficient dynamic memory allocation. In this paper we study two parallel patterns for sliding-window computations and we discuss different implementation variants related to how dynamic memory is managed. The results show that the combined use of an efficient general-purpose memory allocator, and of a custom allocator for the pattern considered, results in significant performance optimizations.