Heterogeneous Execution of Dataflow Programs
Event details
Date | 20.06.2019 |
Hour | 10:30 › 12:30 |
Speaker | Seyedmahyar Emami |
Location | |
Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Babak Falsafi
Thesis advisor: Prof. James Larus
Co-examiner: Prof. Paolo Ienne
Abstract
Economies of scale have shifted computation to the Warehouse-Scale Computers (WSCs) in the cloud. The popularity of certain heavyweight workloads (e.g. machine learning) pushed the cloud providers to integrate accelerators, such as GPUs and FPGAs, to their infrastructure.
The cloud is now comprised of heterogeneous WSCs, combining processing elements of different architectural characteristics. Exploiting the benefits of a heterogeneous WSC is challenging and requires new programming models to efficiently utilize the underlying resources at scale.
Dataflow is a promising model which aids the development of large-scale programs, making it easy for programmers to handle concurrency and to utilize the existing optimized single-machine programs.
Background papers
TensorFlow: A System for Large-Scale Machine Learning, by Abadi, M., et al.
A cloud-scale acceleration architecture, by Caulfield, A. M., et al.
Xilinx Adaptive Compute Acceleration Platform: Versal Architecture, by Gaide, B., et al.
Exam president: Prof. Babak Falsafi
Thesis advisor: Prof. James Larus
Co-examiner: Prof. Paolo Ienne
Abstract
Economies of scale have shifted computation to the Warehouse-Scale Computers (WSCs) in the cloud. The popularity of certain heavyweight workloads (e.g. machine learning) pushed the cloud providers to integrate accelerators, such as GPUs and FPGAs, to their infrastructure.
The cloud is now comprised of heterogeneous WSCs, combining processing elements of different architectural characteristics. Exploiting the benefits of a heterogeneous WSC is challenging and requires new programming models to efficiently utilize the underlying resources at scale.
Dataflow is a promising model which aids the development of large-scale programs, making it easy for programmers to handle concurrency and to utilize the existing optimized single-machine programs.
Background papers
TensorFlow: A System for Large-Scale Machine Learning, by Abadi, M., et al.
A cloud-scale acceleration architecture, by Caulfield, A. M., et al.
Xilinx Adaptive Compute Acceleration Platform: Versal Architecture, by Gaide, B., et al.
Practical information
- General public
- Free