Futures are an attractive way to structure parallel computations. When a thread creates an expression with a keyword future, a new thread is spawned to compute that expression in parallel. When a thread later applies a touch operation to that future, it gets the result of the expression if the result has been computed, and otherwise blocks until the result becomes ready. In this thesis, we explore different aspects of parallel programs with futures, including their theoretical bounds, scheduling, and applications. Researchers have shown futures can have a deleterious effect on cache locality. We will show, however, that if futures are used in a simple, disciplined way, then their negative impact can be much alleviated. This structured use of futures is characteristic of many parallel applications. Futures lend themselves well to dynamic scheduling algorithms, such as work stealing, that ensure high processor utilization. Implementing work stealing on hierarchical platforms, such as NUMA systems and distributed clusters, have recently drawn lots of attention. However, little has been explored on its theoretical bounds. We present lower and upper bounds of work stealing for fork-join programs, a well-studied subclass of parallel-future programs, on hierarchical systems. As originally conceived, a future encapsulates a functional computation without side-effects. Recently, however, futures have been proposed as a way to encapsulate method calls to shared data structures. We propose a new program model, that supports both normal futures without side-effects and linearizable futures that exist for their side-effects. Using this model, we propose the lazy work stealing scheduler that facilitates certain optimizations for linearizable futures and guarantees good time bounds. The processing-in-memory (PIM) model has reemerged recently as a solution to alleviating the growing speed discrepancy between CPU's computation and memory access, commonly known as the memory wall. In this model, some lightweight computing units are directly attached to the main memory, providing fast memory access. We study applications of linearizable futures in the PIM model: operation requests to concurrent data structures are sent as linearizable futures to those computing units to execute. These PIM-managed data structures can outperform state-of-the-art concurrent data structures in the literature.
Liu, Zhiyu,
"Theory and Applications of Parallelism with Futures"
(2017).
Computer Science Theses and Dissertations.
Brown Digital Repository. Brown University Library.
https://doi.org/10.7301/Z0FN14N6
