FPGA-based accelerators are becoming a first-class citizen in data centres. Adding FPGAs in data centres can lead to higher compute densities with improved energy efficiency for latency critical workloads, such as financial applications. However, deployment of FPGAs in datacentres is hindered, as both developers and cloud providers face difficulties.
Application writers need to deal with FPGA interfacing as well as application logic/algorithms. At the other hand, cloud providers are reluctant to deploy FPGAs in large scale, due to the FPGAs lack of sharing that results in reduced utilization and questionable ROI.
To address this problem, VINEYARD introduces VineTalk, a framework that reduces the programming effort associated with FPGA-based accelerators and FPGA virtualization. Specifically in the context of financial applications, VINEYARD has integrated VineTalk with the Xilinx SDAccel development framework and mapped it to the Kintex UltraScale FPGA. Our preliminary evaluation with a use-case of financial applications shows that VineTalk can offer effective FPGA sharing introducing marginal overhead to application execution time. In our evaluation, we use three financial kernels: Black&Scholes, Black-76, and Binomial.
Our results show that VineTalk reduces both programmer effort at the application level by reducing lines of code related to kernel invocation by about 30% with significantly simpler semantics and introduces over- head between 0.9% and 4% compared to native application execution over the FPGA. Finally, VineTalk provides the ability of accelerator sharing from consolidated applications, with less than 2% overhead.
Angelos Bilas (FORTH)