5th Seminar on High Performance Numerical Computing
10h-11h - A Parametric Approach to Steer the Auto-Tuning of Tools in the DOE ACTS Collection, L. A. Drummond – Lawrence Berkeley National Laboratory
The Advanced Computational Software (ACTS) Collection is a set of computational high-performance tools and libraries developed primarily at DOE laboratories. ACTS tools have substantially accelerated the development of high-end computer simulation codes. The collection provides a wide range of scalable numerical algorithms, auto-tuning tools for automatic optimizations of libraries and tools that support the development of parallel codes and their execution. The diversity of the emerging hardware now poses a great challenge for the scalability of the tools. While new programing languages and paradigms are being devised, we look at exploiting a parametric approach to auto-tune ACTS tools and the applications that used them.
Here, we present the ACTS Collection parametric approach to produce highly tuned libraries and scientific applications. A subset of these parameters steer the optimization of the ACTS tools by using a software dependency graph. This graph in turn combines techniques to auto-tune the numerical kernels. The second subset of the parameters is geared for improving the behavior of the numerical kernels in the ACTS tools. These are arguments that are passed by the application to the numerical functionalities in ACTS at run-time. This talk will introduce the ACTS tools functionalities, the ACTS parametric approach auto-tune their installations with some examples and performance results.
11h-11h40 - Co-Array Fortran programming applied to seismic imaging, France Boillod-Cerneux – LIFL/CNRS
† The work is sponsored by TOTAL), with project advisors: Henri Calandra, Terrence Liao** and Saber Feki. (**primary TOTAL contact: firstname.lastname@example.org) and with the most recent result provided by Debjyoti Majumder, Unversity of Houston, and Terrence Liao, TOTAL.
Co-Array Fortran (CAF) is an extension to Fortran language (incorporated in Fortran 2008 Language Standared) for SPMD programming model which allows parallelism. CAF language is part of the PGAS programming model to handle one-sided communication. Its concept is based on the following assumption: ‘A single program is replicated a fixed number of times, each replication having its own set of data objects. Each replication of the program is called an image.’
Since 2009, TOTAL has established collaboration with Prof. Barbara Chapman’s Team at University of Houston, to develop a version of CAF compiler based on OpenUH, an Open64 family of compilers.
The objective of this work is to evaluate the performance of CAF codes using UHCAF compiler vs MPI codes using Intel compiler, on the based code from seismic imaging application for the Oil & Gas exploration.
We concluded that CAF offers more readability and understanding syntax than MPI, regarding the communication operation. Moreover, CAF is better adapted for domain decomposition.
We also studied, based on benchmark communication, the performances of UHCAF compared to MPI for non contiguous data structure, and observed that CAF with GASNet as runtime library offers compatible or even better performance than MPI on non contiguous data structure. In this talk, we will present first the CAF syntax and its concepts. We will then introduce UHCAF operations, and the advantages of CAF regarding the programming ability. We will focus on the UHCAF and Intel MPI performances comparisons for non contiguous data, by using a communication Benchmark. As a conclusion, the performance results obtained from finite difference kernels applied to Isotropic/Tilted Transversely Isotropic media Benchmark, implemented with CAF and MPI are presented.
10h40-12h – Discussions, news and future events
- Inria Saclay - Île-de-France