high performance computing on graphics processing units: hgpu.org

The importance of security infrastructures for high-throughput networks has rapidly grown as a result of expanding internet traffic and increasingly high-bandwidth connections. Intrusion-detection systems (IDSs), such as SNORT, rely upon rule sets designed to alert system administrators of malicious packets. Methods for deep-packet inspection, which often depend upon regular-expression searches, can be accelerated on programmable-logic (PL) architectures using non-deterministic finite automata (NFAs). Prior designs have relied upon register-transfer level (RTL) design descriptions and have achieved efficient resource utilization through fine-grained optimizations. New advances made by field-programmable gate array (FPGA) vendors have led to powerful compiler toolchains for OpenCL and SYCL that allow for rapid development on PL architectures while generating competitive designs in terms of performance. The goal of this work is to evaluate performance differences between a custom, SYCL- and OpenCL-based, acceleration architecture for regular expressions and comparable RTL-based designs. The simplicity of the application, which requires only basic hardware building blocks, adds to the novelty of the comparison. In contrast to prior RTLbased solutions, which show frequency degradation with bandwidth scaling, this approach is able to maintain stable and high operating frequencies at the cost of resource usage. By scaling input bandwidth with multi-character transformations, high-throughput designs can be realized. Using Intel’s OpenCL compiler, throughputs in excess of 17 Gbps can be achieved on Intel’s Arria 10 Programmable Acceleration Card and 19.4 Gbps with Intel’s Stratix 10 Programmable Acceleration Card, outperforming similar designs with RTL, as reported in the literature. SYCL-based designs, synthesized with Intel’s oneAPI compiler show performance degradation but still achieve higher throughput, up to 15.6 Gbps, than past RTL-based implementations. Overall, OpenCL and SYCL development yields both competitive results, when compared to the fine-grained RTL development process, and many ease-of-use improvements and design abstractions.