VWQS: A dispatching mechanism of variable-size tasks in heterogeneous systems

Platforms with accelerators improve both performance and energy efficiency because they are specialized for specific for specific classes of) computations, which leads to-day in high-performance systems that increasingly turn into heterogenous architectures, thus outperforming multicores and GPUs for only a fraction of their power budget. In order to invoke an accelerator, the processor needs to efficiently transfer computation parameters such as operand starting addresses and conf guration data. To address this challenge we present a scalable dispatcher for computational intensive jobs named kernels, of variable complexity. The dispatcher can be easily implemented as a software runtime component or as a decoupled hardware unit, while it features runtime re-programming of weights, of queue size and of decision policies, to support adaptive scheduling, adjusting to application requirements and to system constraints. We evaluated the dispatcher using a simulator with actual kernels runtimes, showing negligible interference to provided scheduling services.