Nowadays, high-performance computing (HPC) not only faces challenges to reach computing performance, it also has to take in consideration the energy consumption. In this context, heterogeneous architectures are expected to tackle this challenge by proposing a mix of HPC and low-power nodes. There is a significant research effort to define methods for exploiting such computing platforms and find a trade-off between computing performance and energy consumption. To this purpose, the topology of the application and the mapping of tasks onto physical resources are of major importance. In this paper we propose an iterative approach based on the exploration of logical topologies and mappings. These solutions are executed onto the heterogeneous platform and evaluated. Based on these results a Pareto front is built, allowing users to select the most relevant configurations of the application according to the current goals and constraints. Experiments have been conducted on a heterogeneous micro-server using a video processing application running on top of a software-distributed shared memory and deployed over a mix of Intel i7 and Arm Cortex A15 processors. Results show that some counterintuitive solutions found by the exploration approach perform better than classical configurations.

Distributed heterogeneous computing systems escalate the problem of choosing the appropriate programming model. Programming models such as message passing are efficient but require low-level management of communications. Higher level of programming such as shared memory are convenient for the application design but they usually have performance issues. With the recent development of distributed heterogeneous systems and new protocols to access remote memories, there is an opportunity for distributed shared memory systems to offer a satisfying level of abstraction while not giving up on performance. In this paper a video processing application is written using MPI, 0MQ and an in-house software-distributed shared memory (S-DSM) backend and deployed over a set of heterogeneous computing boards. Results show that 0MQ implementation is the most efficient but at the price of writing the application with the targeted platform in mind. The S-DSM implementation runs up to 2 times faster than the pure OpenMPI implementation and competes with 0MQ when the data granularity is small.