Continued advances in circuit integration technology has ushered in the era of chip multiprocessor (CMP) architectures as further scaling of the performance of conventional wide-issue superscalar processor architectures remains hard and costly. CMP architectures take advantageof Moore¡¯s Law by integrating more cores in a given chip area rather than a single fastyet larger core. They achieve higher performance with multithreaded workloads. However,CMP architectures pose many new memory hierarchy design and management problems thatmust be addressed. For example, how many cores and how much cache capacity must weintegrate in a single chip to obtain the best throughput possible? Which is more effective,allocating more cache capacity or memory bandwidth to a program?This thesis research develops simple yet powerful analytical models to study two newmemory hierarchy design and resource management problems for CMPs. First, we considerthe chip area allocation problem to maximize the chip throughput. Our model focuses onthe trade-off between the number of cores, cache capacity, and cache management strategies.We find that different cache management schemes demand different area allocation to coresand cache to achieve their maximum performance. Second, we analyze the effect of cachecapacity partitioning on the bandwidth requirement of a given program. Furthermore, ourmodel considers how bandwidth allocation to different co-scheduled programs will affect theindividual programs¡¯ performance. Since the CMP design space is large and simulating only one design point of the designspace under various workloads would be extremely time-consuming, the conventionalsimulation-based research approach quickly becomes ineffective. We anticipate that ouranalytical models will provide practical tools to CMP designers and correctly guide theirdesign efforts at an early design stage. Furthermore, our models will allow them to betterunderstand potentially complex interactions among key design parameters.
展开▼
