Compiling the Functional Data-Parallel Language SaC for Microgrids of Self-Adaptive Virtual Processors

We present preliminary results from compiling the highlevel, functional and data-parallel programming language SaC into a novel multi-core design: Microgrids of Self-Adaptive Virtual Processors (SVPs). The side-effect free nature of SaC in conjunction with its data-parallel foundation make it an ideal candidate for auto-parallelisation. Notwithstanding these favourable pre-conditions, scheduling and data-placement pose major challenges for effective parallelisation of irregular applications because fine-grained dynamic approaches inflict large overheads on conventional architectures. The Microgrid architecture promises a radical shift: thread creation and context switches are implemented in hardware and cause negligible overhead. Likewise, scheduling of tasks to computing resources is catered for by hardware. This paper investigates aspects of the Microgrid architecture which may influence the overall performance of compiled data-parallel programs. In particular, we look at alternative thread creation schemes for ndimensional, data-parallel operations and their effect on overall performance. Furthermore, we investigate the architecture’s capability to cope with workload imbalances within such operations. The paper presents a sequence of experiments on a cycle-accurate emulator of the Microgrid architecture from which we derive some guiding principles for an effective compilation of data parallel operations. Based on these principles, we present a compilation scheme for the data-parallel