Performance Modeling and Scalability Analysis of Stream Computing in ESSPER FPGA Clusters

ESSPER is a flexible and scalable FPGA cluster system designed for reconfigurable HPC, with customizability, scalability, and interoperability with existing HPC systems. The inherent characteristics of ESSPER make it a superior alternative for achieving highly scalable computing solutions, particularly for compute-intensive applications such as numerical simulations and machine learning. However, constraints in communication bandwidth or latency can impact performance and scalability. This paper builds upon a previous performance model study to analyze performance modeling and scalability of streaming computation using a 1D ring of cascaded FPGAs with high-speed, low-latency communication links in ESSPER. For evaluation and verification, we implemented a simple vector addition on ESSPER. The Result indicates that measured performance metrics closely align with predictions from the performance model. This model can also be used to estimate performance for other ESSPER applications.