Resumen de Novel techniques to improve the performance and the energy of vector architectures
The rate of annual data generation grows exponentially. At the same time, there is a high demand to analyze that information quickly. In the past, every processor generation came with a substantial frequency increase, leading to higher application throughput. Nowadays, due to the cease of Dennard scaling, further performance must come from exploiting parallelism.
Vector architectures offer an efficient manner, in terms of performance and energy, of exploting parallelism at data-level by means of instructions that operate over multiple elements at the same time. This is popularly known as Single Instruction Multiple Data (SIMD). Traditionally, vector processors were employed to accelerate applications in research, and they were not industry-oriented. However, vector processors are becoming widely used for data processing in multimedia applications, and entering in new application domains such as machine learning and genomics. In this thesis, we study the circumstances that cause inefficiencies in vector processors, and new hardware/software techniques are proposed to improve the performance and energy consumption of these processors.
We first analyze the behavior of predicated vector instructions in a real machine. We observe that their execution time is dependent on the vector register length and not on the source mask employed. Therefore, a hardware/software mechanism is proposed to alleviate this situation, that will have a higher impact in future processors with wider vector register lengths.
We then study the impact of memory accesses to performance. We identify that an irregular memory access pattern prevents an efficient vectorization, which is automatically discarded by the compiler. For this reason, we propose a near-memory accelerator capable of rearranging data structures and transforming irregular memory accesses to dense ones. This operation may be performed by the devices as the host processor is computing other code regions.
Finally, we observe that many applications with irregular memory access patterns just perform a simple operation on the data before it is evicted back to main memory. In these situations, there is a lack of data access locality, leading to an inefficient use of the memory hierarchy. For this reason, we propose to utilize the accelerators previously described to compute directly near memory.
