Adding data-movement constructs to the PGAS parallel computing model

Authors:
Leonardo Uribe;Pascal Costanza;Theo D'Hondt
Affiliations:
Vrije Universiteit Brussel - VUB, Brussels, Belgium;Vrije Universiteit Brussel - VUB, Brussels, Belgium;Vrije Universiteit Brussel - VUB, Brussels, Belgium
Venue:
Proceedings of the 7th European Lisp Workshop
Year:
2010

Citing 5
Cited 0

The processor-memory bottleneck: problems and solutions

Crossroads - Computer architecture
The Paralation Model: Architecture-Independent Parallel Programming

The Paralation Model: Architecture-Independent Parallel Programming
Experiments with Quadtree Representation of Matrices

ISAAC '88 Proceedings of the International Symposium ISSAC'88 on Symbolic and Algebraic Computation
The potential of the cell processor for scientific computing

Proceedings of the 3rd conference on Computing frontiers
Automatic data movement and computation mapping for multi-level parallel architectures with explicitly managed memories

Proceedings of the 13th ACM SIGPLAN Symposium on Principles and practice of parallel programming

Quantified Score

Hi-index	0.00

Visualization

Abstract

Data movement is a key concept in parallel computing. Given that data movement is generally slower than data processing, it is important to guarantee an efficient use of the communication facilities of the hardware. The MPI communication protocol is the most common model to allocate and move data in High Performance Computing. MPI is a very expressive model but is also a very low-level one. More high-level models such as the PGAS (Partitioned Global Address Space) model allow easier abstraction but limit expressivity. In this paper we present an extension of the PGAS model to include more expressivity while conserving the abstract representations of the model. Specifically, we use the concept of data shape to represent data distribution and use a generalized concept of function to move and process data.