John K. Antonio

Software support for heterogeneous computing

As a result of advances in high-speed digital communications, researchers have begun to use collections of different high-performance machines in concert to execute computationally intensive application tasks. Existing high-performance machines typically achieve only a fraction of their peak performance on certain portions of such application programs; that is, there is a gap between average sustained performance and the machine’s peak performance. One reason for this is that different subtasks of an application can have different computational requirements that are best processed by different types of machine architectures. Thus, an important approach to high-performance computing is to construct a heterogeneous computing (HC) environment, consisting of a variety of machines interconnected by high-speed links, orchestrated to perform an application whose subtasks have diverse execution requirements. In addition to how well a subtask matches a machine, many factors must be considered to exploit optimally the power of an HC suite of machines. These include the time to move data shared by subtasks executed on different machines, the operating system overhead involved in stopping a task on one machine and restarting it on another, the ability to execute subtasks concurrently on all or some subset of the machines in the suite, and the machine and intermachine network load caused by other users of the HC system. There are many instances of successful implementations of application tasks across suites of heterogeneous machines. Typically, however, current users of HC systems must decompose the application task into appropriate subtasks themselves, decide on which machine to execute each subtask, code each subtask specifically for its target machine, and determine the relative execution schedule for the subtasks. The automation of this process is a long-term goal in the field of HC, but research conducted toward this goal should produce tools that will aid the users of HC systems until full automation is possible. Even though the field of HC is relatively new, it is very active. This paper is a brief outline of the software support challenges for HC addressed in Siegel et al. [1996], and readers interested in more details are referred to Eshaghian [1996], Freund and Siegel [1993], Freund and Sunderam [1994], Siegel et al. [1996], and Sunderam [1995]. The first step in using an HC system is to construct the application program. A programming language used in an HC environment must be compilable into efficient code for any machine in the HC suite, and the program specification should facilitate the decomposition of an application task into appropriate subtasks. One model for automated HC consists of four stages: (1) determina-

A fast distributed shortest path algorithm for a class of hierarchically clustered data networks

A distributed algorithm is presented that can be used to solve the single-destination shortest path (SDSP) problem or the all-pairs shortest path (APSP) problem for a class of clustered data networks. The network graph is assumed to be characterized with a balanced hierarchically clustered (BHC) topology. The BHC topology is introduced in this paper and is shown to be a realistic characterization for a large class of interconnected data networks. For certain types of BHC topologies, the SDSP problem can be solved with computation and communication time complexities of O(log n), assuming one processor is available at each of the n number of nodes. Assuming p processors are available at each node, computation and communication time complexities of O((n/p) log n) and O(n log n) are achievable, respectively, for solving the APSP problem. It is also shown that the algorithm converges in an asynchronous environment. >

Optimal trajectory planning for spray coating

The problem of how to optimally traverse a spray applicator around a surface to be coated is formulated as a type of optimization problem known as a constrained variational problem. An optimal trajectory for a spray applicator is defined to be one that results in minimal variation in accumulated film thickness on the surface. For each surface point and for each feasible position and orientation of the applicator, a value for the instantaneous rate of film accumulation is assumed to be known. Empirical data and/or estimates for these values can be readily incorporated in the formulation. By making realistic approximations, the proposed constrained variational problem is transformed into a finite dimensional constrained optimization problem. Numerical studies are included that illustrate the utility of the problem formulation and the effectiveness of applying standard nonlinear programming techniques for determining solutions.<<ETX>>

Fast solution techniques for a class of optimal trajectory planning problems with applications to automated spray coating

Optimal trajectory planning problems are often formulated as constrained variational problems. In general, solutions to variational problems are determined by appropriately discretizing the underlying objective functional and solving the resulting nonlinear differential equation(s) and/or nonlinear programming problem(s) numerically. These general solution techniques often require a significant amount of time to be computed, and therefore are of limited value when optimal trajectories need to be frequently computed and/or recomputed. In this paper, a realistic class of optimal trajectory planning problems is defined for which the existence of fast numerical solution techniques are demonstrated. To illustrate the practicality of this class of trajectory planning problems and the proposed solution techniques, three optimal trajectory planning problems for spray coating applications are formulated and solved. Based on the proposed discretization technique, it is shown that these problems can be reduced to either a linear program or a quadratic program, which are readily solved.

Cost-Minimizing Scheduling of Workflows on a Cloud of Memory Managed Multicore Machines

Workflows are modeled as hierarchically structured directed acyclic graphs in which vertices represent computational tasks, referred to as requests, and edges represent precedent constraints among requests. Associated with each workflow is a deadline that defines the time by which all computations of a workflow should be complete. Workflows are submitted by numerous clients to a scheduler that assigns workflow requests to a cloud of memory managed multicore machines for execution. A cost function is assumed to be associated with each workflow, which maps values of relative workflow tardiness to corresponding cost function values. A novel cost-minimizing scheduling framework is introduced to schedule requests of workflows so as to minimize the sum of cost function values for all workflows. The utility of the proposed scheduler is compared to another previously known scheduling policy.

Complexity of gradient projection method for optimal routing in data networks

In this paper, we derive a time-complexity bound for the gradient projection method for optimal routing in data networks. This result shows that the gradient projection algorithm of Goldstein-Levitin-Poljak type formulated by Bertsekas (1982), Bertsekas and Gallager (1987) and Bertsekas et al. (1984) converges to within /spl epsi/ in relative accuracy in O(/spl epsi//sup -2/h/sub min/N/sub max/) number of iterations, where N/sub max/ is the number of paths sharing the maximally shared link, and h/sub min/ is the diameter of the network. Based on this complexity result, we also show that the one-source-at-a-time update policy has a complexity bound which is O(n) times smaller than that of the all-at-a-time update policy, where n is the number of nodes in the network. The result of this paper argues for constructing networks with low diameter for the purpose of reducing complexity of the network control algorithms. The result also implies that parallelizing the optimal routing algorithm over the network nodes is beneficial.

A Probabilistic Power Prediction Tool for the Xilinx 4000-Series FPGA

The work described here introduces a practical and accurate tool for predicting power consumption for FPGA circuits. The utility of the tool is that it enables FPGA circuit designers to evaluate the power consumption of their designs without resorting to the laborious and expensive empirical approach of instrumenting an FPGA board/chip and taking actual power consumption measurements. Preliminary results of the tool presented here indicate that an error of less than 5% is usually achieved when compared with actual physical measurements of power consumption.

Planning spatial paths for automated spray coating applications

Automated spray coating is an important component of finishing operations in many industries. Robotic manipulators are used in such units to traverse around the object to be coated. This paper describes a framework for determining a good spatial path for the manipulator, in order to achieve uniform deposition over the surface and reduce wastage of coating material outside the surface. By imposing an arc-length based conditioning, the formulated variational problem is shown to be well-posed. The variation in deposition rate, the deposition outside the surface, and the arc length are minimized in a weighted combination. The necessary conditions for a minimizing spatial path are given by a set of nonlinear differential equations that are numerically solved.

Is an alligator better than an armadillo? [interconnection networks]

Multiprocessor interconnection network effectiveness differs considerably across applications and operating environments, but even if these variables were fixed, cost and performance metrics must be chosen. Scientifically determining the best network is as difficult as saying with certainty that one animal is better than another. We explore the problems of determining which metrics or weighted set of metrics designers should use to compare networks and how they should apply these metrics to yield meaningful information. We also look at problems in conducting fair and scientific evaluations.

A highly parallel algorithm for multistage optimization problems and shortest path problems

Abstract It appears that all of the known algorithms for solving multistage optimization problems are based explicitly on standard dynamic programming concepts. Such algorithms are inherently serial in the sense that computation must be completed at the current stage before meaningful computation can begin at the next stage. In this paper we present a technique which recursively divides the original problem into a set of smaller problems which can be solved in parallel. This technique is based on the recursive application of a simple aggregation procedure. For a multistage process with n stages, we show that our new algorithm can achieve a time complexity of O(log n). In contrast, competing algorithms based exclusively on the standard dynamic programming technique can only achieve a time complexity of Φ(n). The new algorithm is designed to operate on a tightly coupled parallel computer. As some important applications, it is shown that our algorithm can serve as a fast and efficient means of decoding convolutional codes, solving shortest path problems, and determining minimum-fuel flight paths.