ARTM: a lightweight fork-join framework for many-core embedded systems

Embedded architectures are moving to multi-core and many-core concepts in order to sustain ever growing computing requirements within complexity and power budgets. Programming many-core architectures not only needs parallel programming skills, but also efficient exploitation of fine grain parallelism at both architecture and runtime levels. Scheduler reactivity is however increasingly important as tasks granularity is reduced, in order to keep the overhead of the scheduling to a minimum. This paper presents a lightweight fork-join framework for scheduling fine grain parallel tasks on embedded many-core systems. The asynchronous nature of the fork-join model used in this framework permits to dramatically decrease its scheduling overhead. Experimentation conducted in this paper show that the overhead induced by this framework is of 33 cycles per scheduled task. Also, we show that near-ideal speedup can be obtained by the ARTM framework for data parallel applications and that ARTM achieves better results than other state of the art parallelization techniques.