Document Actions
Component Integration
Supporting Technologies for Component Integration
The goal of this research is to develop compiler technologies and
library designs that will make it possible to automatically construct
domain-specific development environments for high-performance
applications from collections of components. This effort will develop
advanced compiler technology to integrate collections of components
into a high-performance application without sacrificing the performance
of hand-integrated codes.
In the strategy we envision, programs would use a high-level scripting language such as Matlab or Python to coordinate invocation of library operations, although traditional languages such as Fortran and C++ could also serve this purpose. Scripting languages typically treat library operations as black boxes and thus fail to achieve acceptable performance levels for compute-intensive applications. Previously, researchers have improved performance by translating scripts to a conventional programming language and using whole-program analysis and optimization. Unfortunately, this approach leads to long script compilation times and has no provision to exploit the domain knowledge of library developers.
To address these issues we are pursuing a new approach called “telescoping languages,” in which libraries that provide component operations accessible from scripts are extensively analyzed and optimized in advance. In this scheme, language implementation consists of two phases. The offline translator generation phase digests annotations describing the semantics of library routines, combines them with its own analysis to generate an optimized version of the library, and produces a language translator that understands library entry points as language primitives. The script compilation phase invokes the generated compiler to produce an optimized base language program. The generated compiler must (1) propagate variable property information throughout the script, (2) use a high-level “peephole” optimizer based on library annotations to replace sequences of calls with faster sequences, and (3) select specialized implementations for each library call based on parameter properties at the point of call.
We will use this strategy to attack the problem of making component integration efficient enough to be practical for high-performance scientific codes. Of particular importance in this context is the problem of efficiently integrating data structure components (e.g., sparse matrices) with functional components (e.g., linear algebra). This work will begin with a simple prototype of Matlab (or Python) that includes arrays with data distribution. Specific array distributions for sparse matrices will be explored as a way of understanding the crucial performance issues. In the long term, this may lead to a new strategy for introducing parallelism into Matlab and other scripting languages—by distributing the arrays across multiple processors and performing computations close to the data. (The parallel Matlab effort is leveraged through funding from the NSF ST-HEC effort. In this project we hope to apply this work to ASC codes.)
Once the Matlab array prototype has been explored, we will focus on the Marmot mesh data structures with the goal of demonstrating a prototype with adequate efficiency for use in production codes based on these components. The ultimate goal is to make it possible to quickly substitute different mesh data structures in a code without rewriting the functional components and vice versa.
If this effort is to succeed, it must take into account two important realities. First, many components will be constructed using object-oriented languages, so techniques for optimizing such languages are critical. Second, the execution environments for the resulting programs may be distributed, so the implementation must consider the performance implications of distributed systems, even if the applications are compiled together.
With these considerations in mind, we plan to pursue research in five fundamental directions:
Toolkits for Building Problem-Solving Systems: The effort will focus on the production of tools for defining and building new domain specific PSEs, including:
Advanced Component Integration Systems: This effort will explore the application of telescoping languages technology to the component integration problem, with a particular emphasis on integrating components that support data structures with those that implement functionality. The effort will also consider technologies for optimizing accesses to the component interfaces emerging from the Marmot code development efforts. The long-term goal of this research is to produce a component integration framework that is efficient enough to be accepted by high-performance application developers, such as those in the LANL ASC program.
Design for Efficient Component Integration: This effort will focus on the design and specification of components that can be used in a PSE for high-performance computation. Significant issues will be flexibility and adaptability of the components to both the computations in which they are incorporated and the platforms on which they will be executed. In addition, these components must have architectures that permit the effective management of numerical accuracy. A specific issue of importance is design strategies for efficient data structure components.
Component Systems for Heterogeneous Computing Systems: The key challenge in this area is to construct applications that can be flexibly mapped to heterogeneous computing components and adapt to changes in the execution environment, detecting and correcting performance problems automatically. In this activity, we will explore the meaning of network-aware adaptive component frameworks and what the implementation and optimization challenges are for applications constructed from them. In addition, we will pursue research on middleware to support optimal resource selection in heterogeneous environments. A major byproduct of this work will be performance estimators (described in Section 1.4.1, “Modeling of Application and System Performance”) and mappers that can be used to map applications efficiently to heterogeneous computing systems, such as distributed networks and single-box systems containing different computing components (e.g., vector processors and scalar processors). The latter is a characteristic of several planned HPC computing systems.
Compilation of Object-Oriented Languages: Object-oriented languages like C++, Java, and Python have a number of attractive features for the development of rapid prototyping tools, including full support for software objects, parallel and networking operations, relative language simplicity, type-safety, portability, and a robust commercial marketplace presence leading to a wealth of programmer productivity tools. However, these languages have significant performance problems when used for production applications. In this effort we are studying strategies for the elimination of impediments to performance in object-oriented systems.
To achieve this goal, we must develop new compilation strategies for object-oriented languages such as C++, Java, and Python. This should include interprocedural techniques such as inlining driven by global type analysis and analysis of multithreaded applications. This work would also include new programming support tools for high-performance environments. Initially, this work has focused on Java, through the use of the JaMake high-level Java transformation system developed at Rice in collaboration with the LANL CartaBlanca project. This system includes two novel whole-program optimizations, “class specialization” and “object inlining,” which can improve the performance of high-level, object-oriented, scientific Java programs by up to two orders of magnitude.
In the next phase of research, we will consider how to adapt these strategies to develop tools and compilation strategies that would directly support the code development methodologies to be used in the Marmot effort. Examples include not only the application of object inlining and class specialization, but also the use of type analysis to support the elimination of dynamic dispatch of methods, a major problem for high performance codes written in C++. We will also consider ways to apply these compilation strategies to Python used as a high-level application prototyping system.
In the strategy we envision, programs would use a high-level scripting language such as Matlab or Python to coordinate invocation of library operations, although traditional languages such as Fortran and C++ could also serve this purpose. Scripting languages typically treat library operations as black boxes and thus fail to achieve acceptable performance levels for compute-intensive applications. Previously, researchers have improved performance by translating scripts to a conventional programming language and using whole-program analysis and optimization. Unfortunately, this approach leads to long script compilation times and has no provision to exploit the domain knowledge of library developers.
To address these issues we are pursuing a new approach called “telescoping languages,” in which libraries that provide component operations accessible from scripts are extensively analyzed and optimized in advance. In this scheme, language implementation consists of two phases. The offline translator generation phase digests annotations describing the semantics of library routines, combines them with its own analysis to generate an optimized version of the library, and produces a language translator that understands library entry points as language primitives. The script compilation phase invokes the generated compiler to produce an optimized base language program. The generated compiler must (1) propagate variable property information throughout the script, (2) use a high-level “peephole” optimizer based on library annotations to replace sequences of calls with faster sequences, and (3) select specialized implementations for each library call based on parameter properties at the point of call.
We will use this strategy to attack the problem of making component integration efficient enough to be practical for high-performance scientific codes. Of particular importance in this context is the problem of efficiently integrating data structure components (e.g., sparse matrices) with functional components (e.g., linear algebra). This work will begin with a simple prototype of Matlab (or Python) that includes arrays with data distribution. Specific array distributions for sparse matrices will be explored as a way of understanding the crucial performance issues. In the long term, this may lead to a new strategy for introducing parallelism into Matlab and other scripting languages—by distributing the arrays across multiple processors and performing computations close to the data. (The parallel Matlab effort is leveraged through funding from the NSF ST-HEC effort. In this project we hope to apply this work to ASC codes.)
Once the Matlab array prototype has been explored, we will focus on the Marmot mesh data structures with the goal of demonstrating a prototype with adequate efficiency for use in production codes based on these components. The ultimate goal is to make it possible to quickly substitute different mesh data structures in a code without rewriting the functional components and vice versa.
If this effort is to succeed, it must take into account two important realities. First, many components will be constructed using object-oriented languages, so techniques for optimizing such languages are critical. Second, the execution environments for the resulting programs may be distributed, so the implementation must consider the performance implications of distributed systems, even if the applications are compiled together.
With these considerations in mind, we plan to pursue research in five fundamental directions:
Toolkits for Building Problem-Solving Systems: The effort will focus on the production of tools for defining and building new domain specific PSEs, including:
- Tools for defining and building scripting languages based on well-known platforms, such as Matlab and Python.
- Strategies for scalable parallelization of scripting languages such as Matlab and Python.
- Translation of scripting languages to standard intermediate code, especially languages like C.
- Frameworks for generating optimizers for scripting languages that treat invocations of components from known libraries as primitives in the base language.
- Optimizing translation of intermediate language to distributed and parallel target configurations.
- Assessment of performance/fault tolerance and relation to user code
- Tools for integrating existing code.
- Demonstration of these techniques in specific applications of interest to ASC and LANL, with a special emphasis on codes in the Marmot effort.
Advanced Component Integration Systems: This effort will explore the application of telescoping languages technology to the component integration problem, with a particular emphasis on integrating components that support data structures with those that implement functionality. The effort will also consider technologies for optimizing accesses to the component interfaces emerging from the Marmot code development efforts. The long-term goal of this research is to produce a component integration framework that is efficient enough to be accepted by high-performance application developers, such as those in the LANL ASC program.
Design for Efficient Component Integration: This effort will focus on the design and specification of components that can be used in a PSE for high-performance computation. Significant issues will be flexibility and adaptability of the components to both the computations in which they are incorporated and the platforms on which they will be executed. In addition, these components must have architectures that permit the effective management of numerical accuracy. A specific issue of importance is design strategies for efficient data structure components.
Component Systems for Heterogeneous Computing Systems: The key challenge in this area is to construct applications that can be flexibly mapped to heterogeneous computing components and adapt to changes in the execution environment, detecting and correcting performance problems automatically. In this activity, we will explore the meaning of network-aware adaptive component frameworks and what the implementation and optimization challenges are for applications constructed from them. In addition, we will pursue research on middleware to support optimal resource selection in heterogeneous environments. A major byproduct of this work will be performance estimators (described in Section 1.4.1, “Modeling of Application and System Performance”) and mappers that can be used to map applications efficiently to heterogeneous computing systems, such as distributed networks and single-box systems containing different computing components (e.g., vector processors and scalar processors). The latter is a characteristic of several planned HPC computing systems.
Compilation of Object-Oriented Languages: Object-oriented languages like C++, Java, and Python have a number of attractive features for the development of rapid prototyping tools, including full support for software objects, parallel and networking operations, relative language simplicity, type-safety, portability, and a robust commercial marketplace presence leading to a wealth of programmer productivity tools. However, these languages have significant performance problems when used for production applications. In this effort we are studying strategies for the elimination of impediments to performance in object-oriented systems.
To achieve this goal, we must develop new compilation strategies for object-oriented languages such as C++, Java, and Python. This should include interprocedural techniques such as inlining driven by global type analysis and analysis of multithreaded applications. This work would also include new programming support tools for high-performance environments. Initially, this work has focused on Java, through the use of the JaMake high-level Java transformation system developed at Rice in collaboration with the LANL CartaBlanca project. This system includes two novel whole-program optimizations, “class specialization” and “object inlining,” which can improve the performance of high-level, object-oriented, scientific Java programs by up to two orders of magnitude.
In the next phase of research, we will consider how to adapt these strategies to develop tools and compilation strategies that would directly support the code development methodologies to be used in the Marmot effort. Examples include not only the application of object inlining and class specialization, but also the use of type analysis to support the elimination of dynamic dispatch of methods, a major problem for high performance codes written in C++. We will also consider ways to apply these compilation strategies to Python used as a high-level application prototyping system.
