The Execution Instance Overloading Pattern

Software libraries typically represent substantive investment into a single code base, often capturing the best-known practices, algorithms, and data structures for a given problem domain. Popular libraries become more than a collection of code; they become a language in which programs and algorithms can be expressed. Just as a parallelizing compiler can take program code written in a general-purpose, sequential programming language and produce a parallel program, a parallel library can implement the language of a sequential library to parallelize existing programs. The typical barriers to adoption of a parallel library—unfamiliar interfaces and the need to rewrite a large amount of application code—no longer apply or are lowered substantially. The Execution Instance Overloading pattern facilitates the construction of parallel libraries that inherit the syntax, semantics, and implementation of sequential libraries. Execution Instance Overloading may be applied to algorithms or data structures (collectively called “components”), using three steps: adapt the sequential component with an interface-preserving instance of the Adapter pattern [9, pp. 139–150], overload the sequential component with the parallel one, and optimize cases where the combination of adapter and adaptee would produce suboptimal performance. For instance, consider parallelizing a function that computes the inner product of two vectors and has the following signature: