Optimizing the Utilization of Heterogeneous Systems Using a Single-Source Approach with Dynamic Resource Mapping


Optimizing the Utilization of Heterogeneous Systems Using a Single-Source Approach with Dynamic Resource Mapping

Knodel, O.; Charaf, N.; Stephan, J.; Göhringer, D.

The efficient execution of applications on various heterogeneous resources is an emerging challenge in the heterogeneous landscape of ongoing processing architectures. Especially the mapping of applications to available hardware --regarding to real time requirements, energy parameters, hardware resources and various other system parameters-- is a major academic mission. Based on a single-source C++ software application --only containing SYCL directives-- kernels are generated for different specialized hardware accelerators (e.g. multicore processors, Graphics Processing Units (GPUs) and Field Programmable Gate Arrays (FPGAs)). With the help of hardware emulation the behavior of the execution on different hardware platforms is simulated and a survey is created. While taking additional user objectives and demands (e.g. execution time, hardware resources, power consumption, ...) into account, a virtual Heterogeneous System (vHS) is constructed and then deployed inside a data center or resource-limited edge board. To achieve high system utilization dynamic task offloading between different accelerators and kernel migration is also considered.

Keywords: Heterogeneous Systems; Virtualization; Cloud; Edge Computing; Reconfigurable

  • Poster
    ACACES - Advanced Computer Architecture and Compilation for High-Performance and Embedded Systems, 14.-20.07.2019, Fiuggi, Italy

Permalink: https://www.hzdr.de/publications/Publ-29282
Publ.-Id: 29282