Dynamic Scheduling of AES Cores for Aperiodic Tasks on Multi-tenant Cloud FPGAs

Field Programmable Gate Arrays (FPGAs) have long been utilized in systems benefiting from hardware acceleration of processes unsuitable for execution on a traditional processor. Accordingly, as much of the world pivots from on-site datacenters and computing resources to hybrid or cloud based platforms, Multi-processing System-on-Chip (MPSoC) FPGAs are increasingly being employed in cloud computing systems to speed up many computation-intensive applications. In cloud computing, multi-tenant FPGAs are constantly space- and time-shared among multiple tenants dynamically by leveraging the partial reconfiguration property of FPGAs. With increasing security and privacy concerns introduced by these memory-based volatile devices, most countermeasures rely on encryption and decryption engines such as AES (Advanced Encryption Standard) cores for user data protection. However, their high-resource requirements and long latency limit the number of such engines that can be implemented in hardware. They often become a performance bottleneck during peak time. In this paper, we propose a scheduling algorithm for aperiodic tasks to dynamically share multiple AES cores and hence to improve their utilization and overall system performance. Extensive experimental measurements on an FPGA development board featuring a Xilinx Ultrascale+ FPGA demonstrate the efficacy of our mechanism.