Investigating data throughput and partial dynamic reconfiguration in a commodity FPGA cluster framework

There are many computational kernels where parallelism can be exploited in applica- tion specific hardware, yielding significant speedup over a general purpose processor based solution. Commodity cluster computing technologies have been combined with FPGA co- processors, resulting in even greater performance capability through the exploitation of multiple levels of parallelism. One particularly economic solution both in terms of cost and power consumption is to cluster hybrid FPGAs with commodity network intercon- nects. Hybrid FPGAs combine embedded microprocessors with reconfigurable hardware resources on a single chip offering lower power consumption and cost compared to a tra- ditional I/O bus FPGA coprocessor solution. While there is a lot of promise in using com- modity hybrid FPGAs in a cluster configuration, the design flow and performance char- acteristics of such systems are currently a limiting factor to the range of applications that could benefit from such a system. The contribution of this thesis is a framework for clustering commodity FPGAs which integrates high speed DMA data transfers with a flexible FPGA resource sharing scheme enabled through partial reconfiguration. The framework includes an embedded Linux op- erating system, with a custom device driver to manage data transfers and hardware recon- figuration. User space tools for cluster computing including ssh and MPI are deployed allowing tasks to be split among nodes in the cluster. Performance analysis is performed with a homogeneous cluster composed of four Virtex-5 FXT based FPGA boards. The results demonstrate the advantages over previous work in terms of data throughput and reconfiguration, as well as promote future research efforts