Paul Mackerras

Parallel volume rendering and data coherence

The two key issues in implementing a parallel ray-casting volume renderer are the work distribution and the data distribution. We have implemented such a renderer on the Fujitsu AP1000 using an adaptive image-space subdivision algorithm based on the worker-farm paradigm for the work distribution, and a distributed virtual memory, implemented in software, to provide the data distribution. Measurements show that this scheme works efficiently and effectively utilizes the data coherence that is inherent in volume data.

Exploiting data coherence to improve parallel volume rendering

We have implemented a parallel volume renderer that successfully manages work and data distribution by exploiting data coherence-the tendency of neighboring pixels to use the same data during rendering, particularly when rendering volume data. This flexible, powerful renderer uses ray-casting on a Fujitsu AP1000 to generate high-quality images of volume data sets with other geometrically defined structures, such as a set of coordinate axes or a world map. This article focuses on our schemes for work and data distribution. Using image-space work distribution to partition a 2D image among processing nodes, and distributed virtual memory to assign 3D volume data, this renderer effectively and efficiently parallelizes volume rendering.<<ETX>>

An efficient resource allocation scheme for gang scheduling

Gang scheduling is currently the most popular scheduling scheme for parallel processing in a time shared environment. One major drawback of using gang scheduling is the problem of fragmentation. The conventional method to alleviate this problem is to allow jobs running in multiple time slots. However our experimental results show that simply applying this method alone cannot solve the problem of fragmentation, but on the contrary it may eventually degrade the efficiency of system resource utilisation. In this paper we introduce an efficient resource allocation scheme which effectively incorporates the ideas of re-packing jobs, running jobs in multiple slots and minimising time slots into the buddy allocation system to significantly improve the system and job performance. Because there is no process migration involved in job re-packing, this scheme is particularly suitable for clustered parallel computing systems.

Implementing MPI under AP/Linux

A preliminary MPI library has been implemented for the Fujitsu AP1000+ multicomputer running the AP/Linux operating system. Under this environment, parallel programs may be dedicated to a fixed partition, or a number of parallel programs may share a partition. Therefore, the MPI library has been constructed so that messaging operations can be driven by polling and/or interrupt techniques. It has been found that polling works well when a single parallel program is running on a given partition, and that interrupt-driven communication makes far better use of the machine when multiple parallel programs are executing. Gang scheduling of multiple parallel programs which use polling was found to be relatively ineffective.

Index bit permutations for automatic data redistribution

PISTON is a machine-independent software framework for developing scientific applications on parallel computers. It presents a consistent data-parallel distributed memory model across a wide range of architectures. It has been implemented on MIMD, SIMD and SMP architectures. In this paper, we describe PISTONs implementation of index bit permutations (IBP) as a means of performing automatic regular data redistributions. A theoretical analysis of IBPs is derived and the predicted performance is compared with the actual performance of an IBP implementation on the Fujitsu API000. A detailed examination of the performance of IBPs on two common data redistributions is compared to the performance of hand-coded implementations of the same data redistributions in order to determine the effectiveness of IBPs. Based on this analysis, we generalize to describe what architectural features of a MIMD machine impact on the performance of IBPs and show that they are an efficient means of implementing regular data redistributions on MIMD parallel architectures.

Representation of an image by nonorthogonal basis functions with use of the fast Fourier transform

The importance of nonorthogonal basis sets for representing images is discussed. To find such a representation necessitates diagonalizing overlap matrices that contain billions of elements, even for modest image sizes (256 × 256 pixels). By standard techniques this is completely intractable. An alternative method based on the fast Fourier transform is presented. Not only is it much faster, but it fits well within existing image-processing software.

A planner for time-space coordination of robots in a structured workspace

In industrial applications, tasks for robotic workcells are typically specified by programs written in conventional programming languages such as VAL-II and Karel. Devising planners which synthesize programs from high-level specification statements is a research problem of considerable difficulty, owing to the multiagent, real-time nature of the activities of a robotic workcell. An experimental planner for such high-level task specification is described. The planner is based on abstracting the time-space coordination aspects typical of a wide range of manufacturing tasks. The planner architecture consists of three modules, with the functions of plan generation (providing partially elaborated plans), space-time planning (yielding fully elaborated plans), and the monitoring of plan execution.<<ETX>>