Gossamer: A Lightweight Programming Framework for Multicore Machines

The key to performance improvements in the multicore era is for software to utilize the available concurrency. A recent paper [3] summarizes the challenges and describes twelve so-called dwarfs—types of computing and communication patterns that occur in parallel programs. One of the key points in the paper is that a general programming model has to be able to accommodate all of the patterns defined by the dwarfs, singly or in combination. The challenge is to do so both simply and efficiently. Parallel programming can be supported by providing a threads library [14, 15, 2]; by modifying compilers to extract parallelism or use optimistic code execution techniques [4, 6, 5]; by using concurrency features of existing languages such as Java or C#; by designing new programming languages such as Erlang [16], Fortress [24], X10 [22], and ZPL [7]; or by annotating sequential programs with directives that specify concurrency and synchronization, as in Cilk [11], Cilk++ [8], OpenMP [20], and others [25, 19, 1, 12, 17, 21]. All of these approaches are valuable and are producing useful results, but the last approach—annotating programs—has, in our opinion, the most potential to be simultaneously simple, general, and efficient. In particular, annotations are easier to use than libraries because they hide lots of bookkeeping details, and they are simpler to learn than an entire new programming language. Annotation-based approaches also have efficient implementations. However, no existing approach is general enough to support all the computational patterns (dwarfs) defined in [3]. This paper describes Gossamer, an annotation-based approach that is general as well as simple and efficient. Gossamer has three components: (1) a set of high-level annotations that one adds to a sequential program (C in our case) in order to specify concurrency and synchro-