Heterogeneous Architectures: Versatile Utilization and Programming
- Share
- Partager sur Facebook
- Partager sur LinkedIn
Action Team
The effective programming and operation of heterogeneous architectures require the development of methods and tools for designing systems, whether embedded or not. The HEAVEN action team is part of PERSYVAL-Lab’s PCS research focus area.
Project Description
Computer architectures are becoming increasingly complex, featuring massive parallelism, hierarchically organized memory, and heterogeneous processing units. These architectures are extremely difficult to program because they often force application programmers to choose between portability and performance.
Although standard programming environments such as OpenMP are currently evolving to support the execution of applications on different types of processing units, these approaches suffer from two main problems. First, to leverage heterogeneous processing units at the application level, programmers must explicitly handle hardware-specific low-level mechanisms, such as memory transfers between host memory and a coprocessor’s private memory, for example. Second, as programming environments evolve toward heterogeneous programming, the focus is primarily on CPU/GPU (Central ProcessingUnit/Graphics Processing Unit) platforms, making certain hardware accelerators still difficult to utilize from a general-purpose parallel application.
The FPGA (Field-Programmable Gate Array) is one such example. Unlike CPUs and GPUs, this hardware accelerator can be configured to meet the needs of the application. It contains networks of programmable logic blocks that can be interconnected to build a circuit tailored to the target application. For example, FPGAs can be configured to accelerate portions of code known to perform poorly on CPUs or GPUs. The energy efficiency of FPGAs is also one of the main advantages of this type of accelerator over GPUs, which encourages the scientific community to consider FPGAs as a key component of large-scale, low-power heterogeneous multicore platforms.
However, only a small portion of the community is currently considering programming FPGAs, as the designs must be created using low-level description languages such as VHDL (VHSIC Hardware Description Language), which application programmers are not familiar with.
The main objective of this project isto improve the accessibility of heterogeneous architectures containing FPGA accelerators for parallel application developers. The proposed project focuses on three main aspects:
- Portability: We do not want application developers to have to completely redesign their applications in order to take advantage of FPGA devices. This means that standard parallel programming environments such as OpenMP must be extended to support FPGAs. Improving application portability also means leveraging most of the hardware-specific low-level mechanisms at the runtime level;
- Performance: We want our solution to be flexible enough to make the most of all heterogeneous platforms containing FPGA devices based on specific performance requirements, such as computational throughput or power consumption, for example;
- Experiments: Testing FPGA accelerators on real-world scientific applications is also a key component of our project proposal. In particular, the solutions developed in this project will enable us to compare architectures across real-world applications in various fields, such as signal processing and computational finance.
Project leaders
- olivier.muller
imag.fr (Olivier Muller), Associate Professor at Grenoble INP - UGA, TIMA Laboratory - francois.broquedis
imag.fr (François Broquedis), Associate Professor at Grenoble INP - UGA, LIG Laboratory
Members
- sylvain.huet
gipsa-lab.grenoble-inp.fr (Sylvain Huet), Associate Professor at Grenoble INP - UGA, GIPSA Lab - jose-ernesto.gomez-balderas
gipsa-lab.grenoble-inp.fr (José-Ernesto Gomez-Balderas), associate professor at the University of Georgia (UGA), GIPSA Lab - jerome.lelong
imag.fr (Jérôme Lelong), Associate Professor at Grenoble INP - UGA, LJK Laboratory - frederic.desprez
inria.fr (Frédéric Desprez), research director at INRIA, LIG laboratory - georgios.christodoulis
imag.fr (Georgios Christodoulis), Ph.D. candidate, LIG laboratory - Sébastien Pairault, Master’s 2 intern, GIPSA Lab
- thierry.a.gautier
inria.fr (Thierry Gautier), tenured researcher at INRIA, LIP Laboratory (external collaborator)
News
- [July 10, 2018] Best Paper Award at the ReCoSoC 2018 conference for the paper “An FPGA target for the StarPU heterogeneous runtime system,” by Georgios Christodoulis, Manuel Selva, François Broquedis, Olivier Muller, and Frédéric Desprez.
- [July 3, 2018] Presentation of HEAVEN by Georgios Christodoulis at the LIG conference.
- [06/05/18] Rostyslav Brylov has joined the HEAVEN project as a master’s intern. For three months, he will work under the supervision of Olivier Muller on porting the Rhodinia HPC Benchmark to the HEAVEN platform.
- [March 24, 2018] Best Paper Award at the ATSIP 2018 conference for the paper "Efficient parallelization of the GMM background subtraction algorithm on a multi-core platform for moving object detection," by Lhoussein Mabrouk, Sylvain Huet, Dominique Houzet, Said Belkouch, Abdelkrim Hamzaoui, and Yahya Zennayi
- [02/01/18] Manuel Selva has joined the HEAVEN project as a postdoctoral researcher.
- [July 1, 2017] Lhoussein Mabrouk has joined the HEAVEN project as a doctoral student, with his position partially funded by the project. He will work under the supervision of Sylvain Huet on porting signal processing kernels to heterogeneous platforms.
- [11/03/16] Olivier Muller presented the HEAVEN project at the LIP seminar in Lyon.
- [10/13/16] Sébastien Pairault defended his master’s thesis titled: “Build and Evaluate a CPU/GPU/FPGA Heterogeneous Platform for Embedded Computer Vision Applications.”
- [10/12/16] Olivier Muller gave a talk at the FPGA workshop at UIUC (Urbana-Champaign), which is closely linked to the HEAVEN project.
- [June 29, 2016] Georgios Christodoulis presented the HEAVEN project at the JLESC workshop in Lyon
- [03/01/16] Sébastien Pairault has joined the HEAVEN project as a master’s intern. He will work under the supervision of Sylvain Huet and Ernesto Gomez-Balderas on porting signal processing kernels to heterogeneous platforms incorporating FPGAs.
- [11/24/15] HEAVEN Kickoff Meeting Held at TIMA
-
[10/01/15] Georgios Christodoulis has begun a Ph.D. in the HEAVEN project. Frédéric Desprez is also joining the project as Georgios’s advisor. Georgios will be co-supervised by Frédéric Desprez, Olivier Muller, and François Broquedis.
- Share
- Partager sur Facebook
- Partager sur LinkedIn