Jagannathan Ramanujam - US grants

New techniques for parallelizing scientific computing programs for distributed memory multiprocessors will be developed in this project. In particular, the problem of compiling nested loops will be investigated. The solution of a number of tiling problems will be attempted first. Then examination of methods of determining deadlock-free partitions, size, allocation, and scheduling of tiles and minimizing communication for efficient parallel execution of programs on distributed memory multiprocessors will be carried out. In addition, new tools for optimizations dealing with architectural capabilities of individual processors will be investigated.

The goal of this research is to ease the burden of programming high-performance parallel computers. The project investigates several key issues in this regard: (a) the design of language extensions or new languages for machine-independent parallel programming; (b) algorithms for automatic code and data mapping (both static and dynamic) for scientific programs; (c) code generation, and optimization techniques for locality, inter-processor communication etc.; (d) analysis and optimization of unstructured computations such as those arising in fluid dynamics codes; (e) runtime compilation of irregular sparse matrix codes that exploits both parallelism and data locality; (f) interprocedural analysis and optimization; (g) integration of task and data parallelism; and (h) high-performance input and output.

Jagannathan Ramanujam, Louisiana State University


The performance of programs on modern processors depends critically on
how their memory access characteristics can be matched to the
multi-level memory hierarchy commonly used in these processor
architectures. The goal of this project is derive compiler
transformations to improve the memory performance of scientific
computations. In particular, a combination of program restructuring
and memory layout transformations of data will be derived to handle a
larger class of programming constructs than perfect nests and regular
memory accesses. This project will study several important problems,
including: (a) strategies to integrate tiling and data shackling in
order to effectively orchestrate the movement of data through memory
hierarchies; (b) issues in the design of a sophisticated
locality-enhancing compiler for regular and irregular codes; (c)
extensive experimental evaluation of locality-enhancing
transformations; (d) insights on the interaction between techniques
for exploiting instruction-level parallelism and register-level reuse;
and (e) possible insights on improvements in the design of memory
systems for applications, including the design of application-specific
cache architectures. Most importantly, these compiler techniques will
allow users to easily exploit the enormous computation power in modern
processor architectures.

0103933
Ramanujam, Jagannathan
Louisiana State University & Agricultural and Mechanical College

CISE Postdoctoral Associates in Experimental Computer Science: CISE Postdoctoral Research and Training in Advanced Compiler Optimization

Modern processors extensively use memory hierarchies with multiple levels of caches in order to cope with the widening gap between processor and memory speeds. As a result, the performance of programs depend critically on their memory access characteristics and how these are matched to the memory hierarchy of the processors. While several compiler transformations have been proposed towards enhancing locality, even for programs with regular memory access patterns such as those in dense linear algebra computations, the best compiler-optimized codes do not match the performance of library implementations.

One goal of this research is to train a postdoctoral research associate in the area of developing compiler transformations (either data or computation or a combination of the two) to handle a larger class of programming constructs than perfect nests and regular memory accesses. Initially, the associate will build on data shackling, which has been recently proposed as a data-centric approach to the problem of optimizing locality. To accomplish this goal, the associate will be trained to design an optimization strategy that will integrate data shackling (a data-centric approach) and tiling (a control-centric approach), and implement it in a compiler and evaluate its effectiveness. The associate will:

1) Study a large collection of applications with both regular and irregular memory access patterns, including scientific computing codes and multimedia applications,

2) Design heuristics for improving the effectiveness and applicability of data shackling, develop methods to integrate data shackling and tiling, and implement these in a compiler, and

3) Perform extensive experimental evaluation of the techniques on several benchmarking applications.

Ponnuswamy Sadayappan of Ohio State is supported by the Chemistry Division under the Information Technology Research (ITR) program to develop program synthesis tools that will facilitate high-performance parallel programming for electronic structure calculations. Co-PI's include Gerald Baumgartner and Russ Pitzer of Ohio State, Jagannathan Ramanujam of Louisiana State, and Marcel Nooijen of Princeton, (the latter two via collaborative proposals CHE-0121706 and CHE-0121383). This team of computer scientists and computational chemists will develop a tensor contraction engine that can synthesize efficient parallel code in Fortran or C from an input specification expressed in a high-level notation, for a number of target architectures. This tool will be made freely available to other developers of quantum chemistry software.

The development of high-performance parallel programs for scientific applications is complicated by the effects of algorithm choice on memory access costs and communication overhead. Currently available tools for software development and performance modeling/optimization do not provide adequate support to developers of scientific code. This research will provide a novel approach to the automated synthesis of high-performance parallel programs, with the particular emphasis on electronic structure codes widely employed in chemistry, physics, and materials science.

The goal in this project is the development of new compile/runtime techniques that can enable efficient locality enhanced parallel execution of programs that are conveniently expressed in a high-level programming model. The project will develop performance-model driven approaches to efficient execution of complex applications with multi-level parallelism, with convenient abstractions for developing out-of-core applications, and will investigate applications in a number of scientific domains, focusing on requirements for and challenges in realizing high performance with a global shared view of data. In addition the project will develop compile/runtime techniques to support efficient execution of global-shared-view parallel programs that operate on non-array data structures. The tools and technologies to be developed in this project will target demonstrated needs in various areas of scientific computing. The software products arising from this work will be made freely available for use and extension by others.

Abstract

PROPOSAL NO: 0508245
INSTITUTION: Southern University
PRINCIPAL INVESTIGATOR: Guang-Lin Zhao
TITLE: NER: Integration of Ab-Initio Computation with Large Scale Molecular Dynamics Simulation for Nanomaterials Research

Ab-initio quantum mechanics calculation is a state-of-the-art method in materials research. Complex nanomaterials may involve thousands, even millions, of atoms per unit cell or super-cell. Computations for the complex nanomaterials are beyond the limits of traditional ab-initio quantum calculations. Classical molecular dynamics (MD) simulations, on the other hand, can probe the properties of these systems based on pre-developed interatomic potentials. However, the usefulness of the method is limited by the reliability of the interatomic potential, particularly for complex nanomaterials. The objective of this project is to develop a new computational method and related computer code (computer software) that integrates ab-initio quantum computations with MD simulations. The resulting software will have the capability of MD calculations with the reliability of ab-initio method. The proposed research will have a broad impact on the simulations of nanomaterials for understanding and in some cases for predicting the properties of nanomaterials. Such understanding, based on quantum mechanics at a microscopic level, will shed light on possible mechanism(s) to improve the desired properties of nanomaterials in such a way that it will reduce expensive and redundant experimentation.

This workshop covers aspects of languages, compiler techniques, run-time environments, and architectures for parallel and high-performance computing. Since 1988, the annual LCPC workshops have provided a forum for leading researchers and practitioners to present their latest work and to exchange ideas about future directions. The papers published in the proceedings of the workshop have been widely cited over the years and many of them have represented the first public presentation of innovative ideas. The participation is truly international with speakers from the US, Japan and Europe. While the workshop covers many topics, the emphasis has always been software, e.g., compilers, languages, run-time systems, performance evaluation for parallel systems. The proceesdings of this 18th LCPC Workshop are published by Springer-Verlag, thus making the discussions of this forum available to the broader community.

Background

The difficulty of developing high-performance software using the available languages and tools is being recognized as one of the most significant challenges today in the effective use of high-performance computers. As computers have increased in achievable performance, making it feasible to accurately model more complex phenomena, the time and effort required to develop the software has become the bottleneck in many areas of science and engineering.
This proposal seeks to develop a performance-model driven a compiler optimization framework that integrates the algebraic model for loop representation/transformation with a search-based approach for loop fusion, loop tiling and data/work partitioning. The result of the project will make scientific applications, such as quantum chemistry calculations, and other parallel modeling and simulations, easier programmable, and will cut project implementation and development time.



Intellectual Merit

The goal of this proposal is to develop a framework for compiler optimization that performs loop transformations using performance models such as cache miss cost, disk I/O cost, and inter-processor communication cost, that can be expected to correlate directly with measured performance. Since it will generally be infeasible to analytically determine optimal parameters, or even create cost models that are expressible as algebraic functions of pertinent parameters, our approach is to use search strategies in a potentially large parameter space. This novel optimization framework has potential for high payoffs in generating high-performance code.

Broader Impact

Compiler technology can be very effective in reducing the time for developing applications in
several areas of science and engineering without sacrificing performance. There is an increasing need for automated support that can relieve the burden from users, of low-level details needed to optimize performance. The framework we propose to build will be applicable to a number of high-level language models such as Matlab, Global Arrays, UPC, Co-Array Fortran, ZPL etc. The development will be done in the Open64 framework and the resulting software will be made available to others. It is anticipated that the developed framework will be valuable to researchers in academia and research laboratories. Finally, this proposal includes the development of new courses and the training of two graduate students.

Data protection and recovery have become increasing important as business, education, and government depend more and more on digital information.
Failure events do occur such as virus attacks, user errors, defective software/firmware, hardware faults, and site failures etc that cause data damage. To ensure business continuity and minimize loss, data storage systems need data protection and recovery techniques. However, existing technologies have severe limitations and unable to recover data in many situations. This project aims at studying and understanding how data recovery is done in existing data storage systems, and designing new architectures that will overcome the limitations of existing technologies.

In order to study and understand the existing storage architectures, a new mathematical formulation will be developed to model and analyze capabilities and limitations of the storage architectures. This mathematical model provides a rigorous tool for researchers and practitioners to investigate and understand storage system architectures.
Based on the new mathematical model, a class of new data storage system architectures will be designed that will have the maximum data recoverability. The new storage architectures make it possible for organizations of different sizes to have a cost-effective data storage that provides high data availability and allows quick data recovery upon failures. In addition to the theoretical study, experimental prototypes will be developed and implemented to demonstrate the feasibility, performance, reliability, and data recoverability of newly designed storage architectures. Furthermore, the project includes an education component that advocates a shift of emphasis from CPU-centric computer engineering (CE) curriculum to data-centric CE curriculum. The new curriculum provides CE students with in depth knowledge of data processing, data communication, and data storage.

This proposal will be awarded using funds made available by the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).

In this project, investigators propose to develop a programming environment for easing the development of portable high-performance applications for GPUs and accelerators ? by automatic generation of OpenCL code from annotated C programs provided by the user. The proposed work is motivated by recent advances in polyhedral based approaches for powerful transformations of affine computations that have enabled the development of the Pluto automatic parallelization/optimization system.

The developments will result in enhancement of the widely used gcc compiler providing broad impact across the sciences. The education of students engaged in the work will be central to the project.

The goal of this project is to build a new software infrastructure for
continuing research on compiler optimizations for tensor computations.
Motivated by the successes of the model-driven search-based
optimization approach of the Tensor Contraction Engine (TCE) and the
polyhedral model-based transformations, an optimization infrastructure
is being developed that combines the key aspects of the TCE and the
polyhedral models and provides the flexibility to continue research on
optimizing and automatically parallelizing tensor computations for
parallel and/or distributed computations for any machine architecture,
including multi-cores and GPUs.

This novel optimization framework has potential for high payoffs in
enabling future research and eventually in generating high-performance
code for an important class of quantum chemistry codes and other
tensor computations. After further research, it will become a
valuable tool for reducing the time to develop high-performance
applications in several areas of science and engineering.

Project Shelob is a major effort to support the computational science research, education and training requirements of investigators using modern high performance computing systems equipped with graphical processing units, or GPUs. GPUs are nearly identical to the high-end graphics cards found in personal computers dedicated to gaming applications. The same capabilities that support games with highly realistic real-time animations can be adapted to vastly improve the speed of getting answers on complex science problems. This comes at a cost, however, as the programming becomes more difficult, a burden considering that programming modern parallel computing systems is already a difficult task. The Shelob system provides a dedicated platform to allow experimenting on a production grade system, allowing new methods to be worked out, and providing a platform that can be used for teaching which does not interrupt the critical workflow on other research production systems.
The Shelob system is a compute cluster that allows for parallel programming using distributed memory methods, and adds the ability for nodes to incorporate GPU processing. Such heterogeneous systems require programmers to understand both conventional message-passing (i.e. MPI) programming methods, and the methods specific to GPU programming. The only way researchers can determine how to mix the methods for best performance is to have system-level access to modify and adjust settings as necessary.
There are high expectations for Shelob, not the least of which is instilling excitement in high school and undergraduate students over the possibilities presented for work and research in high performance computing. Programs such as Research Experiences for Undergraduates and the Beowulf Bootcamp, provide unique opportunities to students. Multi-institutional research groups, such as Cactus and Pluto, hope to develop easier ways to make use of the power promised by the Shelob system hardware. Other groups, such as the Louisiana-wide LA-SiGMA project, hope to develop the next generation of codes to support material science research and the development of novel materials for industrial applications.

This is an award to acquire a compute cluster at LSU. The computer is a heterogeneous HPC cluster named SuperMIC containing both Intel Xeon Phi and NVIDIA Kepler K20X GPU (graphics processing unit) accelerators. The intent is to conduct research on programming such clusters while advancing projects that are dependent on HPC. The efforts range from modeling conditions which threaten coastal environments and test mitigation techniques; to simulating the motions of tumors/organs in cancer patients due to respiratory actions to aid radiotherapy planning and management. The burden of learning highly complex hybrid programming models presents an enormous software development crisis and demands a better solution. SuperMIC will serve as the development platform to extend current programming frameworks, such as Cactus, by incorporating GPU and Xeon Phi methods. Such frameworks allow users to move seamlessly from serial to multi-core to distributed parallel platforms without changing their applications, and yet achieve high performance. The SuperMIC project will include training and education at all levels, from a Beowulf boot camp for high school students to more than 20 annual LSU workshops and computational sciences distance learning courses for students at LONI (Louisiana Optical Network Initiative) and LA-SiGMA (Louisiana Alliance for Simulation-Guided Materials Applications) member institutions. These include Southern University, Xavier University, and Grambling State University - all historically black colleges and universities (HBCU) which have large underrepresented minority enrollments. The SuperMIC cluster will be used in the LSU and LA-SiGMA REU and RET programs. It will impact the national HPC community through resources committed to the NSF XSEDE program and the Southeastern Universities Research Association SURAgrid. The SuperMIC will commit 40% of the usage of the machine to the XSEDE XRAC allocation committee.

The cost of data movement through the memory hierarchy is very high in current computer systems, relative to the cost of performing arithmetic operations, both in terms of time and energy. It is expected to become even more dominant in future computer systems. Therefore optimizing data access costs will become increasingly critical in the coming years. This has great impact on algorithm design and implementation on such systems. A characterization of data access complexity is only known today for a few classes of algorithms, such as dense linear algebra, and has required the use of problem-specific analysis techniques. This proposal presents novel ideas for developing automated tools and techniques to characterize the inherent data access complexity of arbitrary algorithms. An automatable general approach for data access complexity of programs will have significant impact on deriving highly effective algorithms and their implementations, understanding the impact of future architectures on algorithm design, and on advances in compilers aimed at improving data access costs of programs.

The proposal develops novel ideas to address a fundamental problem of increasing importance -- developing tools and techniques for characterization of the inherent data access complexity of algorithms. The work develops the following: (i) the first unifying approach for developing both upper and lower bounds for the inherent data access complexity of a CDAG under a common underlying theoretical model; (ii) an approach to remove a fundamental current limitation in the use of the popular reuse distance analysis metric; (iii) a scalable and versatile dynamic analysis infrastructure of use to application developers, compiler writers and architecture designers.

The U.S. Economic Development Administration, through its Investing in Manufacturing Communities Partnership (IMCP) initiative, recently named the entire east-west, 200-mile span of South Louisiana centered around Baton Rouge (BR) as one of the nation's strategic manufacturing regions. This "chemical corridor" is the home of hundreds of chemical manufacturing facilities and refineries worth billions of dollars. Despite the major economic significance of this region, the BR area suffers from critical problems including i) crippling transportation issues, and (ii) high levels of crime. Heavy traffic congestion is one of the main reasons why manufacturing industry in BR is reluctant to build new plants and hire more people in this region; it is also a key factor in making the region unattractive for new investments from elsewhere in the U.S. The violent crime rate is substantially higher than that of similarly- sized regions, further affecting economic development in this important region. The goal of this S&CC planning proposal is to build a partnership between community stakeholders and a multidisciplinary team of academic researchers. By considering economic and social issues in a holistic manner, this partnership will develop research concepts that will promote and employ S&CC technologies to help stakeholders tackle the major factors affecting the region's economic progress. Through multidisciplinary team- building and strong community engagement, a proposed integrative S&CC research concept will have significant impacts on various disciplines and communities for planning and developing smarter cities. In particular, the research concept will be directly aligned with the strategic plans determined by the city's Smart City Committee and Subcommittees. Through the proposed Web portal, PIs will share the developed outcomes and information collected from the project with researchers from other higher education including Southern University , a local HBCU. The most significant impact will be seen through quality of life measures pre- and post-project implementations. Building community-wide tools to help stakeholders of all missions addressing crime- and traffic-related challenges for all citizens in the region has real impact on quality of life and economic health, making the region more attractive for growth.


From a technical perspective, this project proposes to build a multidisciplinary research team by integrating different research groups with significant research strength in the areas of High Performance Computing (LSU CCT), Big Data Analysis and Cybersecurity (Computer Science), Sensors (Electrical Engineering), Violence Prevention (Social Work and Sociology), and Transportation (Civil Engineering). The intellectual goal is to define challenging problems and develop research concepts via offline workshops, tutorials, and an online Web portal, which allows an easy access to integrative cyberinfrastructure for computing, storage, and software tools. The software tools we plan to develop will enable predictive analyses on heterogeneous data collected from city infrastructure, public open data from the city of Baton Rouge, and relevant social network data. The developed research concepts will increase the research capacity of every individual research group, enhance the understanding of cross-disciplinary demands, and advance state-of-art technologies for the design of smart and connected communities.