Lars Bergstrom

Nested data-parallelism on the gpu

Graphics processing units (GPUs) provide both memory bandwidth and arithmetic performance far greater than that available on CPUs but, because of their Single-Instruction-Multiple-Data (SIMD) architecture, they are hard to program. Most of the programs ported to GPUs thus far use traditional data-level parallelism, performing only operations that operate uniformly over vectors. NESL is a first-order functional language that was designed to allow programmers to write irregular-parallel programs - such as parallel divide-and-conquer algorithms - for wide-vector parallel computers. This paper presents our port of the NESL implementation to work on GPUs and provides empirical evidence that nested data-parallelism (NDP) on GPUs significantly outperforms CPU-based implementations and matches or beats newer GPU languages that support only flat parallelism. While our performance does not match that of hand-tuned CUDA programs, we argue that the notational conciseness of NESL is worth the loss in performance. This work provides the first language implementation that directly supports NDP on a GPU.

Garbage collection for multicore NUMA machines

Modern high-end machines feature multiple processor packages, each of which contains multiple independent cores and integrated memory controllers connected directly to dedicated physical RAM. These packages are connected via a shared bus, creating a system with a heterogeneous memory hierarchy. Since this shared bus has less bandwidth than the sum of the links to memory, aggregate memory bandwidth is higher when parallel threads all access memory local to their processor package than when they access memory attached to a remote package. This bandwidth limitation has traditionally limited the scalability of modern functional language implementations, which seldom scale well past 8 cores, even on small benchmarks. This work presents a garbage collector integrated with our strict, parallel functional language implementation, Manticore, and shows that it scales effectively on both a 48-core AMD Opteron machine and a 32-core Intel Xeon machine.

Data-only flattening for nested data parallelism

Data parallelism has proven to be an effective technique for high-level programming of a certain class of parallel applications, but it is not well suited to irregular parallel computations. Blelloch and others proposed nested data parallelism (NDP) as a language mechanism for programming irregular parallel applications in a declarative data-parallel style. The key to this approach is a compiler transformation that flattens the NDP computation and data structures into a form that can be executed efficiently on a wide-vector SIMD architecture. Unfortunately, this technique is ill suited to execution on todays multicore machines. We present a new technique, called data-only flattening, for the compilation of NDP, which is suitable for multicore architectures. Data-only flattening transforms nested data structures in order to expose programs to various optimizations while leaving control structures intact. We present a formal semantics of data-only flattening in a core language with a rewriting system. We demonstrate the effectiveness of this technique in the Parallel ML implementation and we report encouraging experimental results across various benchmark applications.

Lazy tree splitting

Nested data-parallelism (NDP) is a language mechanism that supports programming irregular parallel applications in a declarative style. In this paper, we describe the implementation of NDP in Parallel ML (PML), which is a part of the Manticore system. One of the main challenges of implementing NDP is managing the parallel decomposition of work. If we have too many small chunks of work, the overhead will be too high, but if we do not have enough chunks of work, processors will be idle. Recently, the technique of Lazy Binary Splitting was proposed to address this problem for nested parallel loops over flat arrays. We have adapted this technique to our implementation of NDP, which uses binary trees to represent parallel arrays. This new technique, which we call Lazy Tree Splitting (LTS), has the key advantage of performance robustness, i.e., it does not require tuning to get the best performance for each program. We describe the implementation of the standard NDP operations using LTS and present experimental data that demonstrate the scalability of LTS across a range of benchmarks.

Arity raising in manticore

Compilers for polymorphic languages are required to treat values in programs in an abstract and generic way at the source level. The challenges of optimizing the boxing of raw values, flattening of argument tuples, and raising the arity of functions that handle complex structures to reduce memory usage are old ones, but take on newfound import with processors that have twice as many registers. We present a novel strategy that uses both control-flow and type information to provide an arity raising implementation addressing these problems. This strategy is conservative -- no matter the execution path, the transformed program will not perform extra operations.

Engineering the servo web browser engine using Rust

All modern web browsers - Internet Explorer, Firefox, Chrome, Opera, and Safari - have a core rendering engine written in C++. This language choice was made because it affords the systems programmer complete control of the underlying hardware features and memory in use, and it provides a transparent compilation model. Unfortunately, this language is complex (especially to new contributors!), challenging to write correct parallel code in, and highly susceptible to memory safety issues that potentially lead to security holes. Servo is a project started at Mozilla Research to build a new web browser engine that preserves the capabilities of these other browser engines but also both takes advantage of the recent trends in parallel hardware and is more memory-safe. We use a new language, Rust, that provides us a similar level of control of the underlying system to C++ but which statically prevents many memory safety issues and provides direct support for parallelism and concurrency. In this paper, we show how a language with an advanced type system can address many of the most common security issues and software engineering challenges in other browser engines, while still producing code that has the same performance and memory profile. This language is also quite accessible to new open source contributors and employees, even those without a background in C++ or systems programming. We also outline several pitfalls encountered along the way and describe some potential areas for future improvement.

Programming in Manticore, a heterogenous parallel functional language

The Manticore project is an effort to design and implement a new functional language for parallel programming. Unlike many earlier parallel languages, Manticore is a heterogeneous language that supports parallelism at multiple levels. Specifically, the Manticore language combines Concurrent ML-style explicit concurrency with fine-grain, implicitly threaded, parallel constructs. These lectures will introduce the Manticore language and explore a variety of programs written to take advantage of heterogeneous parallelism. At the explicit-concurrency level, Manticore supports the creation of distinct threads of control and the coordination of threads through first-class synchronous-message passing. Message-passing synchronization, in contrast to shared-memory synchronization, fits naturally with the functional-programming paradigm. At the implicit-parallelism level, Manticore supports a diverse collection of parallel constructs for different granularities of work. Many of these constructs are inspired by common functional-programming idioms. In addition to describing the basic mechanisms, we will present a number of useful programming techniques that are enabled by these mechanisms.