Pipelined fission for stream programs with dynamic selectivity and partitioned state
Date
Authors
Editor(s)
Advisor
Supervisor
Co-Advisor
Co-Supervisor
Instructor
Source Title
Print ISSN
Electronic ISSN
Publisher
Volume
Issue
Pages
Language
Type
Journal Title
Journal ISSN
Volume Title
Usage Stats
views
downloads
Series
Abstract
There is an ever increasing rate of digital information available in the form of online data streams. In many application domains, high throughput processing of such data is a critical requirement for keeping up with the soaring input rates. Data stream processing is a computational paradigm that aims at addressing this challenge by processing data streams in an on-the-fly manner, in contrast to the more traditional and less efficient store-and-then process approach. In this paper, we study the problem of automatically parallelizing data stream processing applications in order to improve throughput. The parallelization is automatic in the sense that stream programs are written sequentially by the application developers and are parallelized by the system. We adopt the asynchronous data flow model for our work, which is typical in Data Stream Processing Systems (DSPS), where operators often have dynamic selectivity and are stateful. We solve the problem of pipelined fission, in which the original sequential program is parallelized by taking advantage of both pipeline parallelism and data parallelism at the same time. Our pipelined fission solution supports partitioned stateful data parallelism with dynamic selectivity and is designed for shared-memory multi-core machines. We first develop a cost-based formulation that enables us to express pipelined fission as an optimization problem. The bruteforce solution of this problem takes a long time for moderately sized stream programs. Accordingly, we develop a heuristic algorithm that can quickly, but approximately, solve the pipelined fission problem. We provide an extensive evaluation studying the performance of our pipelined fission solution, including simulations as well as experiments with an industrial-strength DSPS. Our results show good scalability for applications that contain sufficient parallelism, as well as close to optimal performance for the heuristic pipelined fission algorithm.