X10 for high-performance scientific computing

Abstract

High performance computing is a key technology that enables large-scale physical simulation in modern science. While great advances have been made in methods and algorithms for scientific computing, the most commonly used programming models encourage a fragmented view of computation that maps poorly to the underlying computer architecture. Scientific applications typically manifest physical locality, which means that interactions between entities or events that are nearby in space or time are stronger than more distant interactions. Linear-scaling methods exploit physical locality by approximating distant interactions, to reduce computational complexity so that cost is proportional to system size. In these methods, the computation required for each portion of the system is different depending on that portion’s contribution to the overall result. To support productive development, application programmers need programming models that cleanly map aspects of the physical system being simulated to the underlying computer architecture while also supporting the irregular workloads that arise from the fragmentation of a physical system. X10 is a new programming language for high-performance computing that uses the asynchronous partitioned global address space (APGAS) model, which combines explicit representation of locality with asynchronous task parallelism. This thesis argues that the X10 language is well suited to expressing the algorithmic properties of locality and irregular parallelism that are common to many methods for physical simulation. The work reported in this thesis was part of a co-design effort involving researchers at IBM and ANU in which two significant computational chemistry codes were developed in X10, with an aim to improve the expressiveness and performance of the language. The first is a Hartree–Fock electronic structure code, implemented using the novel Resolution of the Coulomb Operator approach. The second evaluates electrostatic interactions between point charges, using either the smooth particle mesh Ewald method or the fast multipole method, with the latter used to simulate ion interactions in a Fourier Transform Ion Cyclotron Resonance mass spectrometer. We compare the performance of both X10 applications to state-of-the-art software packages written in other languages. This thesis presents improvements to the X10 language and runtime libraries for managing and visualizing the data locality of parallel tasks, communication using active messages, and efficient implementation of distributed arrays. We evaluate these improvements in the context of computational chemistry application examples. This work demonstrates that X10 can achieve performance comparable to established programming languages when running on a single core. More importantly, X10 programs can achieve high parallel efficiency on a multithreaded architecture, given a divide-and-conquer pattern parallel tasks and appropriate use of worker-local data. For distributed memory architectures, X10 supports the use of active messages to construct local, asynchronous communication patterns which outperform global, synchronous patterns. Although point-to-point active messages may be implemented efficiently, productive application development also requires collective communications; more work is required to integrate both forms of communication in the X10 language. The exploitation of locality is the key insight in both linear-scaling methods and the APGAS programming model; their combination represents an attractive opportunity for future co-design efforts.