TMFab: A Transactional Memory Fabric for Chip Multiprocessors

With the performance of single-core processors approaching its limits, an increased amount of research effort is focused on chip multiprocessors (CMP). However, existing lock-based synchronization methods that are critical to performing parallel computation possess limited scalability and are inherently complex to use while programming. This thesis uses the concept of transactional memory implemented within a synthesizable fabric named TMFab, containing all the requisite hardware components needed to prototype a scalable chip-multiprocessor. Its processor independent nature enables the instantiation and use of any suitable soft-processor core inside the fabric without significant modifications to the fabric hardware. Additionally, the fabric offers scalability on account of its 3D interconnect architecture that supports die-stacking to add additional processor cores to the CMP without increasing its area footprint. The hardware transactional memory system of the fabric reduces performance overheads of transactional operations, allowing transactions to complete execution faster. TMFab is shown to provide speed up as high as 3.44x for correctly partitioned independent transactions and can be used to analyze the points of contention for conflicting transactions. The fabric was synthesized for both Field Programmable Gate Array (FPGA) as well as 90nm semi-custom targets.