Cache coherent commutative operations

Abstract

This thesis presents Coup, a technique that reduces the cost of updates in shared memory systems. In particular, it describes a new cache coherence protocol, MEUSI, and evaluates its performance under simulation in zsim. MEUSI extends the MESI protocol to allow data to be cached in a new update-only state, reducing both block-level thrashing and on-chip network traffic under many parallel workflows. Coup permits both single-word and multi-word commutative data operations, which are implemented as x86-64 ISA extensions. To evaluate single-word instructions, this thesis presents a case study of a new reference counting scheme, and for multi-word commutative operations, this thesis describes the design of a commutative memory allocator. Coup and MEUSI confer significant benefits to the reference counting scheme and the memory allocator, both in terms of performance and ease of programming.