Data-Driven Scenario-Based Application Mapping for Heterogeneous Many-Core Systems

For applications whose workload and execution behavior significantly varies with the input, a single mapping of application tasks to a given target architecture is insufficient. A single mapping may deliver a high-quality solution for the average case but rarely exploits the specific execution behavior of concurrent tasks triggered by each input tuple. E.g., tasks with higher computational demands under certain input should be mapped onto high-performance resources of the heterogeneous architecture. This necessitates mappings that are specialized for specific input data. Yet, due to the large size of input combinations, determining a separate optimized mapping for each individual input workload is not feasible for most applications. As a remedy, we propose to group input data with similar execution characteristics into a selected, small number of so-called workload scenarios for which we supply optimized mappings. In this paper, we provide a data-driven approach for detecting workload scenarios and exploring scenario-optimized mappings based on a collection of input data. The identification of scenarios and the determination of optimized mappings are interdependent: For the data-driven identification of workload scenarios, we have to measure the profiles when executing the application with the given input data for different application mappings. However, to come up with scenario-optimized application mappings, the workload scenarios have to be known. We tackle this interdependence problem by proposing a cyclic design methodology that optimizes both aspects in an iterative fashion. It is shown that with our approach, the latency of two exemplary applications, a ray tracing as well as an image stitching application, can be significantly improved compared to methods that ignore workload scenarios or do not perform the proposed iterative refinement. Furthermore, we demonstrate that our proposal can be used in the context of a hybrid application mapping methodology.