Content area
The high-performance computing market has entered the Petaflop era, and is aiming to the Exascale supercomputer. One of the important approaches to achieving the Exascale performance is the heterogeneous computing, whereby the capability of each node is extended with the addition of various types of computational accelerators, such as Cell, GPUs, and FPGAs. In such heterogeneous (or hybrid) systems, CPUs offer generality over a wide range of applications while specialized accelerators provide better power efficiency and performance-per-dollar for specific computation patterns. However the parallel of the algorithms on the heterogeneous system is difficult. It is different from the traditional architecture. This paper presents two important parallel programming models which could leverage the computing capacity of the heterogeneous system. One is the CPU-Centred model and the other is Accelerator-Centred model. These models are very useful for the current popular GPU-accelerated systems and the emerging CPU and GPU fusion systems.