Abstract
The paper overviews the state of the art in design and implementation of data parallel scientific applications on heterogeneous platforms. It covers both traditional approaches originally designed for clusters of heterogeneous workstations and the most recent methods developed in the context of modern multicore and multi-accelerator heterogeneous platforms.
Highlights
High performance computing systems become increasingly heterogeneous and hierarchical
It was shown that even the static distribution based on simplistic performance models improves the performance of traditional dynamic scheduling techniques by up to 250% [44]. In this overview we focus on parallel scientific applications, where computational workload is directly proportional to the size of data, and dedicated HPC platforms, where: (i) the performance of the application is stable in time and is not affected by varying system load; (ii) there is a significant overhead associated with data migration between computing devices; (iii) optimized architecture-specific libraries implementing the same kernels may be available for different computing devices
The CPM can be a sufficiently accurate approximation of the performance of heterogeneous processors executing a data parallel application if: (i) the processors are general-purpose and execute the same code, (ii) the local tasks are small enough to fit in the main memory but large enough not to fully fit in the processor cache
Summary
High performance computing systems become increasingly heterogeneous and hierarchical. It was shown that even the static distribution based on simplistic performance models (single values specifying the maximum performance of a dominant computational kernel on CPUs and GPUs) improves the performance of traditional dynamic scheduling techniques by up to 250% [44] In this overview we focus on parallel scientific applications, where computational workload is directly proportional to the size of data, and dedicated HPC platforms, where: (i) the performance of the application is stable in time and is not affected by varying system load; (ii) there is a significant overhead associated with data migration between computing devices; (iii) optimized architecture-specific libraries implementing the same kernels may be available for different computing devices.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
