Accessibility navigation


Scalable parallelization of stencils using MODA

Jumah, N. and Kunkel, J. (2019) Scalable parallelization of stencils using MODA. In: P^3MA workshop, pp. 142-154.

[img]
Preview
Text - Accepted Version
· Please see our End User Agreement before downloading.

427kB

It is advisable to refer to the publisher's version if you intend to cite from this work. See Guidance on citing.

To link to this item DOI: 10.1007/978-3-030-34356-9_13

Abstract/Summary

The natural and the design limitations of the evolution of processors, e.g., frequency scaling and memory bandwidth bottlenecks, push towards scaling applications on multiple-node configurations besides to exploiting the power of each single node. This introduced new challenges to porting applications to the new infrastructure, especially with the heterogeneous environments. Domain decomposition and handling the resulting necessary communication is not a trivial task. Parallelizing code automatically cannot be decided by tools in general as a result of the semantics of the general-purpose languages. To allow scientists to avoid such problems, we introduce the Memory-Oblivious Data Access (MODA) technique, and use it to scale code to configurations ranging from a single node to multiple nodes, supporting different architectures, without requiring changes in the source code of the application. We present a technique to automatically identify necessary communication based on higher-level semantics. The extracted information enables tools to generate code that handles the communication. A prototype is developed to implement the techniques and used to evaluate the approach. The results show the effectiveness of using the techniques to scale code on multi-core processors and on GPU based machines. Comparing the ratios of the achieved GFLOPS to the number of nodes in each run, and repeating that on different numbers of nodes shows that the achieved scaling efficiency is around 100%. This was repeated with up to 100 nodes. An exception to this is the single-node configuration using a GPU, in which no communication is needed, and hence, no data movement between GPU and host memory is needed, which yields higher GFLOPS.

Item Type:Conference or Workshop Item (Paper)
Refereed:Yes
Divisions:Faculty of Science > School of Mathematical, Physical and Computational Sciences > Department of Computer Science
ID Code:87577

Downloads

Downloads per month over past year

University Staff: Request a correction | Centaur Editors: Update this record

Page navigation