Nathaniel Nystrom

Firepile: run-time compilation for GPUs in scala

Recent advances have enabled GPUs to be used as general-purpose parallel processors on commodity hardware for little cost. However, the ability to program these devices has not kept up with their performance. The programming model for GPUs has a number of restrictions that make it difficult to program. For example, software running on the GPU cannot perform dynamic memory allocation, requiring the programmer to pre-allocate all memory the GPU might use. To achieve good performance, GPU programmers must also be aware of how data is moved between host and GPU memory and between the different levels of the GPU memory hierarchy. We describe Firepile, a library for GPU programming in Scala. The library enables a subset of Scala to be executed on the GPU. Code trees can be created from run-time function values, which can then be analyzed and transformed to generate GPU code. A key property of this mechanism is that it is modular: unlike with other meta-programming constructs, the use of code trees need not be exposed in the library interface. Code trees are general and can be used by library writers in other application domains. Our experiments show Firepile users can achieve performance comparable to C code targeted to the GPU with shorter, simpler, and easier-to-understand code.

Data-centric Consistency Policies: A Programming Model for Distributed Applications with Tunable Consistency

The consistency level of operations over replicated data is an important parameter in distributed applications. It impacts correctness, performance, and availability. It is now common to find single applications using many different consistency levels at the same time; however, current commercial frameworks do not provide high-level abstractions for specifying or reasoning about different consistency properties of an application. Research frameworks that do tend to require a substantial effort from developers to specify operation dependencies, orderings and invariants to be preserved. We propose an approach for specifying consistency properties based on the observation that correctness criteria and invariants are a property of data, not operations. Hence, it is reasonable to define the consistency properties required to enforce various data invariants on the data itself rather than on the operations. The result is a system that is simpler to describe and reason about. In this paper, we outline an abstract model of programming language constructs and a static checker for data-centric consistency control, and demonstrate this model through a simple prototype programming language implementation.

The Truth, The Whole Truth, and Nothing But the Truth: A Pragmatic Guide to Assessing Empirical Evaluations

An unsound claim can misdirect a field, encouraging the pursuit of unworthy ideas and the abandonment of promising ideas. An inadequate description of a claim can make it difficult to reason about the claim, for example, to determine whether the claim is sound. Many practitioners will acknowledge the threat of unsound claims or inadequate descriptions of claims to their field. We believe that this situation is exacerbated, and even encouraged, by the lack of a systematic approach to exploring, exposing, and addressing the source of unsound claims and poor exposition. This article proposes a framework that identifies three sins of reasoning that lead to unsound claims and two sins of exposition that lead to poorly described claims and evaluations. Sins of exposition obfuscate the objective of determining whether or not a claim is sound, while sins of reasoning lead directly to unsound claims. Our framework provides practitioners with a principled way of critiquing the integrity of their own work and the work of others. We hope that this will help individuals conduct better science and encourage a cultural shift in our research community to identify and promulgate sound claims.

A Scala framework for supercompilation

Supercompilation is a program transformation technique that attempts to evaluate programs as much as possible at compile time. Supercompilation has been used for theorem proving, function inversion, and most notably optimization, especially of functional programs. However, the technique has numerous practical problems that prevent it from being applied in mainstream compilers. In this paper, we describe a framework that can be used for experimenting with supercompilation techniques. Our framework allows supercompilers to be constructed directly from an interpreter. The user specifies the interpreter using rewrite rules and the framework handles termination checking, generalization, and residualization. We demonstrate the approach by implementing a supercompiler for JavaScript.

Harmless compiler plugins

Languages such as Java and Scala allow programmers to write compiler extensions, or plugins, that extend the host programming language with new functionality to enable additional static checking and code transformations. However, by permitting arbitrary code transformations, compiler plugins can change the host language semantics in unexpected ways. Moreover, plugins do not compose. Plugins can interfere with each other such that one plugin can undo the effects of another, or worse, cause another plugin to generate incorrect code. In this paper, we develop a theoretical framework for harmless compiler plugins. Host language programs are annotated to limit the scope of plugins. Plugins may change the termination behavior of code outside these scopes, but they are prohibited from changing the values computed by the original computation. The framework is based on an extension of Welterweight Java and uses an information-flow type system to limit plugin effects.

Implementing resource-aware safe assembly for kernel probes as a dependently-typed DSL

We present construction of resource-aware safe typed assembly language as an EDSL in dependently-typed Idris language. We use this assembly language to compile Linux kernel probes — small pieces of instrumentation code injected directly into the kernel and thus having to satisfy strict safety properties. We believe that the techniques presented can be generally applied to embedding a typed assembly language into a functional language with dependent types.

Tapir: A Language for Verified OS Kernel Probes

Kernel probes allow code to be inserted into a running operating system kernel to gather information for debugging or profiling. Inserting code into the kernel raises a number of safety issues. Current solutions follow one of the two paths: a VM-based approach, where safety properties are checked dynamically by an interpreter, or a static-analysis approach, where probe code is guaranteed to be safe statically. While more attractive, existing static solutions depend on ad-hoc and error-prone analysis. We propose to explore enforcing safety properties using a type system, thus building our analysis on top of the well-studied ground of type theory.

Tapir: a language for verified OS kernel probes

Kernel probes allow code to be inserted into a running operating system kernel to gather information for debugging or profiling. Inserting code into the kernel raises a number of safety issues. Current solutions follow one of the two paths: a VM-based approach, where safety properties are checked dynamically by an interpreter, or a static-analysis approach, where probe code is guaranteed to be safe statically. While more attractive, existing static solutions depend on ad-hoc and error-prone analysis. We propose to explore enforcing safety properties using a type system, thus building our analysis on top of the well-studied ground of type theory.

A Refactoring Library for Scala Compiler Extensions

Compiler plugins enable languages to be extended with new functionality by adding compiler passes that perform additional static checking, code generation, or code transformations. However, compiler plugins are often difficult to build. A plugin can perform arbitrary code transformations, easily allowing a developer to generate incorrect code. Moreover, the base compiler assumes many complex, sometimes undocumented invariants, requiring plugin developers to acquire intimate knowledge of the design and implementation of the compiler. To address these issues in the context of the Scala compiler plugin framework, we introduce Piuma. Piuma is a library that provides, first, an API to perform many common refactoring tasks needed by plugin writers, and, second, a DSL to eliminate much of the boilerplate code required for plugin development. We demonstrate the usefulness of our library by implementing five diverse compiler plugins. We show that, using Piuma, plugins require less code and are easier to understand than plugins developed using the base Scala compiler plugin API.