Towards Efficient Design Space Exploration of FPGA-based Accelerators for Streaming HPC Applications (Abstract Only)

Streaming HPC applications are data intensive and have widespread use in various fields (e.g., Computational Fluid Dynamics and Bioinformatics). These applications consist of different processing kernels where each kernel performs a specific computation on its input data. The objective of the optimization process is to maximize performance. FPGAs show great promise for accelerating streaming applications because of their low power consumption combined with high theoretical compute capabilities. However, mapping an HPC application to a reconfigurable fabric is a challenging task. The challenge is exacerbated by need to temporally partition computational kernels when application requirements exceed resource availability. In this poster, we present work towards a novel design methodology for exploring design space of streaming HPC applications on FPGAs. We assume that the designer can represent the target application with a Synchronous Data Flow Graph (SDFG). In the SDFG, the nodes are compute kernels and the edges signify data flow between kernels. The designer should also determine the problem size of the application and the volume of raw data on each memory source of the SDFG. The output of our method is a set of FPGA configurations that each contain one or more SDFG nodes. The methodology consists of three main steps. In Step 1, we enumerate the valid partitions and the base configurations. In Step 2, we find the feasible base configurations given the hardware resources available and a library of processing kernel implementations. Finally, we use a performance model to calculate the execution time of each partition in Step 3. Our current assumption is that it is advantageous to represent SDFG at a coarse granularity since this enables exhaustive exploration of the design space for practical applications. This approach has yielded promising preliminary results. In one case, the temporal configuration selected by our methodology outperformed the direct mapping by 3X.