Parallelization

Compiling for parallelism is a longstanding topic of compiler research. This book describes the fundamental principles of compiling "regular" numerical programs for parallelism. We begin with an explanation of analyses that allow a compiler to understand the interaction of data reads and writes in different statements and loop iterations during program execution. These analyses include dependence analysis, use-def analysis and pointer analysis. Next, we describe how the results of these analyses are used to enable transformations that make loops more amenable to parallelization, and discuss transformations that expose parallelism to target shared memory multicore and vector processors. We then discuss some problems that arise when parallelizing programs for execution on distributed memory machines. Finally, we conclude with an overview of solving Diophantine equations and suggestions for further readings in the topics of this book to enable the interested reader to delve deeper into the field. Table of Contents: Introduction and overview / Dependence analysis, dependence graphs and alias analysis / Program parallelization / Transformations to modify and eliminate dependences / Transformation of iterative and recursive constructs / Compiling for distributed memory machines / Solving Diophantine equations / A guide to further reading

Distributed-memory multiprocessing systems (DMS), such as Intel's hypercubes, the Paragon, Thinking Machine's CM-5, and the Meiko Computing Surface, have rapidly gained user acceptance and promise to deliver the computing power required to solve the grand challenge problems of Science and Engineering. These machines are relatively inexpensive to build, and are potentially scalable to large numbers of processors. However, they are difficult to program: the non-uniformity of the memory which makes local accesses much faster than the transfer of non-local data via message-passing operations implies that the locality of algorithms must be exploited in order to achieve acceptable performance. The management of data, with the twin goals of both spreading the computational workload and minimizing the delays caused when a processor has to wait for non-local data, becomes of paramount importance. When a code is parallelized by hand, the programmer must distribute the program's work and data to the processors which will execute it. One of the common approaches to do so makes use of the regularity of most numerical computations. This is the so-called Single Program Multiple Data (SPMD) or data parallel model of computation. With this method, the data arrays in the original program are each distributed to the processors, establishing an ownership relation, and computations defining a data item are performed by the processors owning the data.

This book presents a solution to the challenging issue of optimizing expensive-to-evaluate industrial problems such as the hyperparameter tuning of machine learning models. The approach combines two well-established concepts, Surrogate-Based Optimization (SBO) and parallelization, to efficiently search for optimal parameter setups with as few function evaluations as possible. Through in-depth analysis, the need for parallel SBO solvers is emphasized, and it is demonstrated that they outperform model-free algorithms in scenarios with a low evaluation budget. The SBO approach helps practitioners save significant amounts of time and resources in hyperparameter tuning as well as other optimization projects. As a highlight, a novel framework for objectively comparing the efficiency of parallel SBO algorithms is introduced, enabling practitioners to evaluate and select the most effective approach for their specific use case. Based on practical examples, decision support is delivered, detailing which parts of industrial optimization projects can be parallelized and how to prioritize which parts to parallelize first. By following the framework, practitioners can make informed decisions about how to allocate resources and optimize their models efficiently.

This book introduces new compilation techniques, using the polyhedron model for the resource-adaptive parallel execution of loop programs on massively parallel processor arrays. The authors show how to compute optimal symbolic assignments and parallel schedules of loop iterations at compile time, for cases where the number of available cores becomes known only at runtime. The compile/runtime symbolic parallelization approach the authors describe reduces significantly the runtime overhead, compared to dynamic or just‐in-time compilation. The new, on‐demand fault‐tolerant loop processing approach described in this book protects loop nests for parallel execution against soft errors.

Readership This book is devoted to the study of compiler transformations that are needed to expose the parallelism hiddenin a program. This book is notan introductory book to parallel processing, nor is it an introductory book to parallelizing compilers. Weassume thatreaders are familiar withthebooks High Performance Compilers for Parallel Computingby Wolfe [121] and Super­ compilers for Parallel and Vector Computers by Zima and Chapman [125], and that they want to know more about scheduling transformations. In this book we describe both task graph scheduling and loop nest scheduling. Taskgraphschedulingaims atexecuting tasks linked by prece­ dence constraints; it is a run-time activity. Loop nest scheduling aims at ex­ ecutingstatementinstances linked bydata dependences;it is a compile-time activity. We are mostly interested in loop nestscheduling,butwe also deal with task graph scheduling for two main reasons: (i) Beautiful algorithms and heuristics have been reported in the literature recently; and (ii) Several graphscheduling, like list scheduling, are the basis techniques used in task ofthe loop transformations implemented in loop nest scheduling. As for loop nest scheduling our goal is to capture in a single place the fantastic developments of the last decade or so. Dozens of loop trans­ formations have been introduced (loop interchange, skewing, fusion, dis­ tribution, etc.) before a unifying theory emerged. The theory builds upon the pioneering papers of Karp, Miller, and Winograd [65] and of Lam­ port [75], and it relies on sophisticated mathematical tools (unimodular transformations, parametric integer linear programming, Hermite decom­ position, Smithdecomposition, etc.).

This book contains the main results of the German project POPINDA. It surveys the state of the art of industrial aerodynamic design simulations on parallel systems. POPINDA is an acronym for Portable Parallelization of Industrial Aerodynamic Applications. This project started in late 1993. The research and development work invested in POPINDA corresponds to about 12 scientists working full-time for the three and a half years of the project. POPINDA was funded by the German Federal Ministry for Education, Science, Research and Technology (BMBF). The central goals of POPINDA were to unify and parallelize the block-structured aerodynamic flow codes of the German aircraft industry and to develop new algorithmic approaches to improve the efficiency and robustness of these programs. The philosophy behind these goals is that challenging and important numerical appli cations such as the prediction of the 3D viscous flow around full aircraft in aerodynamic design can only be carried out successfully if the benefits of modern fast numerical solvers and parallel high performance computers are combined. This combination is a "conditio sine qua non" if more complex applications such as aerodynamic design optimization or fluid structure interaction problems have to be solved. When being solved in a standard industrial aerodynamic design process, such more complex applications even require a substantial further reduction of computing times. Parallel and vector computers on the one side and innovative numerical algorithms such as multigrid on the other have enabled impressive improvements in scientific computing in the last 15 years.

This compact yet thorough tutorial is the perfect introduction to the basic concepts of solving partial differential equations (PDEs) using parallel numerical methods. In just eight short chapters, the authors provide readers with enough basic knowledge of PDEs, discretization methods, solution techniques, parallel computers, parallel programming, and the run-time behavior of parallel algorithms to allow them to understand, develop, and implement parallel PDE solvers. Examples throughout the book are intentionally kept simple so that the parallelization strategies are not dominated by technical details.

This volume presents a collection of peer-reviewed, scientific articles from the 15th International Conference on Information Technology – New Generations, held at Las Vegas. The collection addresses critical areas of Machine Learning, Networking and Wireless Communications, Cybersecurity, Data Mining, Software Engineering, High Performance Computing Architectures, Computer Vision, Health, Bioinformatics, and Education.