Memory abstractions for parallel programming

Abstract

A memory abstraction is an abstraction layer between the program execution and the memory that provides a different "view" of a memory location depending on the execution context in which the memory access is made. Properly designed memory abstractions help ease the task of parallel programming by mitigating the complexity of synchronization or admitting more efficient use of resources. This dissertation describes five memory abstractions for parallel programming: (i) cactus stacks that interoperate with linear stacks, (ii) efficient reducers, (iii) reducer arrays, (iv) ownershipaware transactions, and (v) location-based memory fences. To demonstrate the utility of memory abstractions, my collaborators and I developed Cilk-M, a dynamically multithreaded concurrency platform which embodies the first three memory abstractions. Many dynamic multithreaded concurrency platforms incorporate cactus stacks to support multiple stack views for all the active children simultaneously. The use of cactus stacks, albeit essential, forces concurrency platforms to trade off between performance, memory consumption, and interoperability with serial code due to its incompatibility with linear stacks. This dissertation proposes a new strategy to build a cactus stack using thread-local memory mapping (or TLMM), which enables Cilk-M to satisfy all three criteria simultaneously. A reducer hyperobject allows different branches of a dynamic multithreaded program to maintain coordinated local views of the same nonlocal variable. With reducers, one can use nonlocal variables in a parallel computation without restructuring the code or introducing races. This dissertation introduces memory-mapped reducers, which admits a much more efficient access compared to existing implementations. When used in large quantity, reducers incur unnecessarily high overhead in execution time and space consumption. This dissertation describes support for reducer arrays, which offers the same functionality as an array of reducers with significantly less overhead. Transactional memory is a high-level synchronization mechanism, designed to be easier to use and more composable than fine-grain locking. This dissertation presents ownership-aware transactions, the first transactional memory design that provides provable safety guarantees for "opennested" transactions. On architectures that implement memory models weaker than sequential consistency, programs communicating via shared memory must employ memory-fences to ensure correct execution. This dissertation examines the concept of location-based memoryfences, which unlike traditional memory fences, incurs latency only when synchronization is necessary.