OpenCL HLS Based Design of FPGA Accelerators for Cryptographic Primitives

Abstract

Modern data centers are being transformed to meet the increased processing needs of specialized workloads with an advantageous total cost of ownership. To this end, the modular design of current microservers allows the inclusion of heterogeneous computing platforms and accelerators to enhance the performance of specific workloads, while improving the power consumption and maintenance costs of the whole system. One of the fundamental application domains for datacenters is represented by bulk data encryption and decryption, as it has to be performed on the data being stored as well as on data being transmitted or received. In this paper we investigate the OpenCL programming practices to realize high-performance FPGA accelerators, thus providing a viable and more versatile alternative to the use of ad-hoc cryptographic accelerators, which are currently available in high-end server CPUs only. We validate our analysis employing AES-128 as our case study, and report energy efficiency improvements of 22.78× with respect to pure software implementations of ISO standard block ciphers.

Accelerating Automotive Analytics: The M2DC Appliance Approach

“…For this purpose, fast disk encryption software support has been explored using GPGPUs [2]. Cryptography SEE proposed in M2DC [7] exploits the OpenCL programming model to realize high-performance FPGA accelerators. This way provides a viable and more versatile alternative to the use of ad-hoc cryptographic accelerators, currently available in high-end server CPUs only.…”

“…In particular, we found out that single work-item, single work-group implementations are more suitable for FPGAs. We also identified other best-practices related to loop optimization, caching techniques, memory coalescing, I/O latency hiding and host-side synchronization, whose details can be found in [7].…”

“…These approaches imply that each process is a separate instance of the R interpreter without any shared memory. These approaches include RMPI 6 , an implementation of MPI bindings in R which enables master-slave like scheduling, and Snow Extension (Simple Network Of Workstations) 7 , which provides a further level of abstraction above existing communication frameworks, masking communications details and enabling the creation of different kinds of clusters. Although they partially address the scalability issues of R, these approaches do not solve the overhead introduced by the interpretation layer and hardly target heterogeneous systems.…”