O-structures: semantics for versioned memory

Abstract

This paper introduces O-structures, a novel architectural memory element that can be used to facilitate parallelism in task-based execution models. Much like register renaming, each write to an O-structure creates a new version of program memory at that location. These versions can be accessed concurrently and out of program order. O-structures provide a set of semantics that match the needs of task-based execution models, specifically allowing tasks to synchronize on specific versions of memory as well as coordinate access when the necessary version is not known at compile time. In this work, we describe O-structures and provide their complete semantics. We also discuss how a task-based execution of basic data structure manipulations on common data structures (arrays, lists, trees, etc) operate. Results are presented that measure the exposed memory-level parallelism (MLP) in these operations. We find that for previously difficult to parallelize data-structures, such as linked lists, binary trees and sparse-matrix codes we see significant memory level parallelism (50--100 operations per cycle) when using O-structures.