Lightweight Parallel Foundations: a model-compliant communication layer

TL;DR

The paper introduces LPF, a communication layer that ensures model compliance and interoperability, enabling the development of immortal algorithms like FFT and PageRank that perform efficiently across diverse environments.

Contribution

LPF provides a set of primitives with performance guarantees, facilitating the creation of portable, optimal algorithms that can be integrated into various parallel and sequential systems.

Findings

LPF FFT matches Intel MKL performance and outperforms FFTW.

LPF enables seamless integration of algorithms into different platforms.

Model compliance does not compromise computational performance.

Abstract

We present the Lightweight Parallel Foundations (LPF), an interoperable and model-compliant communication layer adhering to a strict performance model of parallel computations. LPF consists of twelve primitives, each with strict performance guarantees, two of which enable interoperability. We argue that the principles of interoperability and model compliance suffice for the practical use of immortal algorithms: algorithms that are proven optimal once, and valid forever. These are ideally also implemented once, and usable from a wide range of sequential and parallel environments. This paradigm is evaluated by implementing an immortal fast Fourier transform (FFT) using LPF, and compared to state-of-the-art FFT implementations. We find it performs on par to Intel MKL FFT while consistently outperforming FFTW, thus showing model compliance can be achieved without sacrificing performance. Interoperability encourages the propagation of immortal algorithms as widely as possible. We evaluate this by integrating an LPF PageRank into Spark, without changing any PageRank nor Spark source codes, and while requiring only a minimal interface layer.