Vector support for multicore processors with major emphasis on configurable multiprocessors
Department of Electrical and Computer Engineering
Doctor of Philosophy
Gerbessiotis, Alexandros V.
It recently became increasingly difficult to build higher speed uniprocessor chips because of performance degradation and high power consumption. The quadratically increasing circuit complexity forbade the exploration of more instruction-level parallelism (JLP). To continue raising the performance, processor designers then focused on thread-level parallelism (TLP) to realize a new architecture design paradigm. Multicore processor design is the result of this trend. It has proven quite capable in performance increase and provides new opportunities in power management and system scalability. But current multicore processors do not provide powerful vector architecture support which could yield significant speedups for array operations while maintaining arealpower efficiency.
This dissertation proposes and presents the realization of an FPGA-based prototype of a multicore architecture with a shared vector unit (MCwSV). FPGA stands for Filed-Programmable Gate Array. The idea is that rather than improving only scalar or TLP performance, some hardware budget could be used to realize a vector unit to greatly speedup applications abundant in data-level parallelism (DLP). To be realistic, limited by the parallelism in the application itself and by the compiler's vectorizing abilities, most of the general-purpose programs can only be partially vectorized. Thus, for efficient resource usage, one vector unit should be shared by several scalar processors. This approach could also keep the overall budget within acceptable limits. We suggest that this type of vector-unit sharing be established in future multicore chips.
The design, implementation and evaluation of an MCwSV system with two scalar processors and a shared vector unit are presented for FPGA prototyping. The MicroBlaze processor, which is a commercial IP (Intellectual Property) core from Xilinx, is used as the scalar processor; in the experiments the vector unit is connected to a pair of MicroBlaze processors through standard bus interfaces. The overall system is organized in a decoupled and multi-banked structure. This organization provides substantial system scalability and better vector performance. For a given area budget, benchmarks from several areas show that the MCwSV system can provide significant performance increase as compared to a multicore system without a vector unit.
However, a MCwSV system with two MicroBlazes and a shared vector unit is not always an optimized system configuration for various applications with different percentages of vectorization. On the other hand, the MCwSV framework was designed for easy scalability to potentially incorporate various numbers of scalar/vector units and various function units. Also, the flexibility inherent to FPGAs can aid the task of matching target applications. These benefits can be taken into account to create optimized MCwSV systems for various applications. So the work eventually focused on building an architecture design framework incorporating performance and resource management for application-specific MCwSV (AS-MCwSV) systems. For embedded system design, resource usage, power consumption and execution latency are three metrics to be used in design tradeoffs. The product of these metrics is used here to choose the MCwSV system with the smallest value.
njit-etd2008-072 (105 pages ~ 4,935 KB pdf)
Please complete this Feedback Form to inform us about your experience using this website. It will assist us in better serving your information needs in the future. Thank You!
Created October 9, 2008