Scalable embedded computing through reconfigurable hardware: Comparing DF-Threads, cilk, openmpi and jump

• Published in: Microprocessors and microsystems Vol. 63; pp. 66 - 74
• Main Author: Giorgi, Roberto
• Format: Journal Article
• Language: English
• Published: Kidlington Elsevier B.V 01.11.2018; Elsevier BV
• Subjects: Computer peripherals; Computer programming; Computer simulation; Configurations; Cyber-physical systems; Distributed memory; Distributed shared memory; Embedded systems; FPGA Programming; Memory model; Message passing; Nodes; Performance evaluation; Programming model; Reconfigurable hardware; Reconfigurable systems; System on chip; Transceivers; Workload; Performance evaluation; Reconfigurable systems; Cyber-physical systems; Programming model; Memory model; FPGA Programming; Distributed shared memory
• ISSN: 0141-9331; 1872-9436
• DOI: 10.1016/j.micpro.2018.08.005

Data-Flow Threads (DF-Threads) is a new execution model that permits to seamlessly distribute the workload across several cores (in a multi-core) and several nodes (in a multi-node/multi-board configuration). In this paper, the advance in deploying this execution model is shown while developing it by using a combination of a simulator model (i.e., the COTSon framework) and a reconfigurable hardware platform (i.e., the AXIOM-board). The AXIOM platform consists of a custom board based on the Xilinx Zynq Ultrascale+ ZU9EG, which incorporates the largest FPGA available on that System-on-Chip at the moment, four 64-bit ARM cores and two 32-bit ARM cores, up to 32GiB of main memory and several 16Gbit/s transceivers. While a complete DF-Threads system is still under development, but is already capable of running a full Linux OS and simple applications, so some initial results are presented here. In particular, well-known programming models that are used to exploit the Thread-Level Parallelism such as Cilk, OpenMPI and Jump are compared with DF-thread execution. Cilk is good for multi-cores, but it is not suitable for multi-nodes systems. In the latter cases, the distribution of the workload could be managed partly by the programmer when using programming models such as message-passing (OpenMPI has been chosen for reference) or distributed shared-memory (Jump in our case). The obtained results show that a DF-Thread execution on a cluster of eight 4-core boards can provide a speed-up of more than 14x compared to the same configuration when using OpenMPI and more than 80x when compared with a OpenMPI single core, single node execution.