Record Details

Automatic Optimization of Geometric Multigrid Methods using a DSL Approach

Electronic Theses of Indian Institute of Science

View Archive Info
 
 
Field Value
 
Title Automatic Optimization of Geometric Multigrid Methods using a DSL Approach
 
Creator Vasista, Vinay V
 
Subject Geometric Multigrid Methods
Automatic Optimization
Geometric Multigrid Method
Compiler Optimizations
PolyMage Compiler
DSL Approach
Automatic Optimization
Domain-specific Language Approach
GMG Algorithm
Polyhedral Optimization
Storage Optimization
Computer Science
 
Description Geometric Multigrid (GMG) methods are widely used in numerical analysis to accelerate the convergence of partial differential equations solvers using a hierarchy of grid discretizations. These solvers find plenty of applications in various fields in engineering and scientific domains, where solving PDEs is of fundamental importance. Using multigrid methods, the pace at which the solvers arrive at the solution can be improved at an algorithmic level. With the advance in modern computer architecture, solving problems with higher complexity and sizes is feasible - this is also the case with multigrid methods. However, since hardware support alone cannot achieve high performance in execution time, there is a need for good software that help programmers in doing so.
Multiple grid sizes and recursive expression of multigrid cycles make the task of manual program optimization tedious and error-prone. A high-level language that aids domain experts to quickly express complex algorithms in a compact way using dedicated constructs for multigrid methods and with good optimization support is thus valuable. Typical computation patterns in a GMG algorithm includes stencils, point-wise accesses, restriction and interpolation of a grid. These computations can be optimized for performance on modern architectures using standard parallelization and locality enhancement techniques.
Several past works have addressed the problem of automatic optimizations of computations in various scientific domains using a domain-specific language (DSL) approach. A DSL is a language with features to express domain-specific computations and compiler support to enable optimizations specific to these computations. Halide and PolyMage are two of the recent works in this direction, that aim to optimize image processing pipelines. Many computations like upsampling and downsampling an image are similar to interpolation and restriction in geometric multigrid methods.
In this thesis, we demonstrate how high performance can be achieved on GMG algorithms written in the PolyMage domain-specific language with new optimizations we added to the compiler. We also discuss the implementation of non-trivial optimizations, on PolyMage compiler, necessary to achieve high parallel performance for multigrid methods on modern architectures. We realize these goals by:
• introducing multigrid domain-specific constructs to minimize the verbosity of the algorithm specification;
• storage remapping to reduce the memory footprint of the program and improve cache locality exploitation;
• mitigating execution time spent in data handling operations like memory allocation and freeing, using a pool of memory, across multiple multigrid cycles; and
• incorporating other well-known techniques to leverage performance, like exploiting multi-dimensional parallelism and minimizing the lifetime of storage buffers.
We evaluate our optimizations on a modern multicore system using five different benchmarks varying in multigrid cycle structure, complexity and size, for two-and three-dimensional data grids. Experimental results show that our optimizations:
• improve performance of existing PolyMage optimizer by 1.31x;
• are better than straight-forward parallel and vector implementations by 3.2x;
• are better than hand-optimized versions in conjunction with optimizations by Pluto, a state-of-the-art polyhedral source-to-source optimizer, by 1.23x; and
• achieve up to 1.5$\times$ speedup over NAS MG benchmark from the NAS Parallel Benchmarks.
(The speedup numbers are Geometric means over all benchmarks)
 
Contributor Uday Kumar Reddy, B
 
Date 2018-06-14T06:57:24Z
2018-06-14T06:57:24Z
2018-06-14
2017
 
Type Thesis
 
Identifier http://etd.iisc.ernet.in/2005/3707
http://etd.iisc.ernet.in/abstracts/4577/G28295-Abs.pdf
 
Language en_US
 
Relation G28295