Selecting Locking Primitives for Parallel Programs

The only reason to parallelize a program is to gain performance. However, the synchronization primitives used by parallel programs can consume execessive memory bandwidths, can be subject to memory laten-cies, consume excessive memory, and result in unfair access or even starvation. These problems can overwhelm the performance beneets of parallel execution. Therefore, it is necessary to understand these performance implications of synchronization primitives in addition to their correctness, liveness, and safety properties. This paper presents a pattern language to assist you in selecting synchronization primitives for parallel programs. This pattern language assumes you have already chosen a locking design, perhaps by using a locking design pattern language McK96]. 1 Overview A lock-based parallel program uses synchronization primitives to deene critical sections of code in which only one CPU or thread may execute concurrently. For example, Figure 1 presents a fragment of parallel code to search and update a linear list. In this C-code example, the lt next eld links the individual elements together, the lt key eld contains the search key, and the lt data eld contains the data corresponding to that key. The section of code between the S LOCK() and the S UNLOCK() primitives is a critical section. Only one CPU at a time may be executing in this critical section. A poor choice of locking primitive can result in excessive overhead and poor performance under heavy load. The pattern language in this paper will help you determine what kind of locking primitive to use. This paper considers a few straightforward test-and-set, queued, and reader/writer locks, which will handle most situations. This paper presents the implementation level counterpart to a locking design pattern language McK96]. Section 2 therefore gives an overview of locking design patterns. Section 3 describes the forces common to all of the patterns. Section 4 overviews contexts in which these patterns are useful. Section 5 presents several indexes to the patterns. Section 6 presents the patterns themselves. Although design and implementation are often treated as separate activities, they are almost always deeply intertwined. Therefore, this section presents a brief overview of design-level patterns and the forces that act on them. 2.1 Overview of Locking Design Patterns This paper refers to the following locking design patterns: Sequential Program: A design with no parallelism, ooering none of the beneets or problems associated with parallel programs. Code Locking: A design where locks are associated with speciic sections of …