Date of Award
Doctor of Philosophy in Computer Engineering - (Ph.D.)
Electrical and Computer Engineering
Alexandros V. Gerbessiotis
Vector Processors (VPs) created the breakthroughs needed for the emergence of computational science many years ago. All commercial computing architectures on the market today contain some form of vector or SIMD processing.
Many high-performance and embedded applications, often dealing with streams of data, cannot efficiently utilize dedicated vector processors for various reasons: limited percentage of sustained vector code due to substantial flow control; inherent small parallelism or the frequent involvement of operating system tasks; varying vector length across applications or within a single application; data dependencies within short sequences of instructions, a problem further exacerbated without loop unrolling or other compiler optimization techniques. Additionally, existing rigid SIMD architectures cannot tolerate efficiently dynamic application environments with many cores that may require the runtime adjustment of assigned vector resources in order to operate at desired energy/performance levels.
To simultaneously alleviate these drawbacks of rigid lane-based VP architectures, while also releasing on-chip real estate for other important design choices, the first part of this research proposes three architectural contexts for the implementation of a shared vector coprocessor in multicore processors. Sharing an expensive resource among multiple cores increases the efficiency of the functional units and the overall system throughput. The second part of the dissertation regards the evaluation and characterization of the three proposed shared vector architectures from the performance and power perspectives on an FPGA (Field-Programmable Gate Array) prototype. The third part of this work introduces performance and power estimation models based on observations deduced from the experimental results. The results show the opportunity to adaptively adjust the number of vector lanes assigned to individual cores or processing threads in order to minimize various energy-performance metrics on modern vector- capable multicore processors that run applications with dynamic workloads. Therefore, the fourth part of this research focuses on the development of a fine-to-coarse grain power management technique and a relevant adaptive hardware/software infrastructure which dynamically adjusts the assigned VP resources (number of vector lanes) in order to minimize the energy consumption for applications with dynamic workloads. In order to remove the inherent limitations imposed by FPGA technologies, the fifth part of this work consists of implementing an ASIC (Application Specific Integrated Circuit) version of the shared VP towards precise performance-energy studies involving high- performance vector processing in multicore environments.
Beldianu, Spiridon Florin, "Vector coprocessor sharing techniques for multicores: performance and energy gains" (2012). Dissertations. 326.