Scheduling Workflow Applications Based on Multi-source Parallel Data Retrieval in Distributed Computing Networks

摘要

Many scientific experiments are carried out in collaboration with researchers around the world to use existing infrastructures and conduct experiments at massive scale. Data produced by such experiments are thus replicated and cached at multiple geographic locations. This gives rise to new challenges when selecting distributed data and compute resources so that the execution of applications is time- and cost-efficient. Existing heuristic techniques select 'best' data source for retrieving data to a compute resource and subsequently process task-resource assignment. However, this approach of scheduling, which is based only on single source data retrieval, may not give time-efficient schedules when: (ⅰ) tasks are interdependent on data, (ⅱ) the average size of data processed by most tasks is large and (ⅲ) data transfer time exceeds task computation time by at least one order of magnitude. In order to address these characteristics of data-intensive applications, we propose to leverage the presence of replicated data sources, retrieve data in parallel from multiple locations and thus achieve time-efficient schedules. In this article, we propose two multi-source data-retrieval-based scheduling heuristic that assigns interdependent tasks to compute resources based on both data retrieval time and task-computation time. We carry out experiments using real applications and deploy them on emulated as well as real environments. With a combination of data retrieval and task-resource mapping technique, we show that our heuristic produces time-efficient schedules that are better than existing heuristic-based techniques for scheduling application workflows.