Aske Plaat

MagPIe: MPI's collective communication operations for clustered wide area systems

Writing parallel applications for computational grids is a challenging task. To achieve good performance, algorithms designed for local area networks must be adapted to the differences in link speeds. An important class of algorithms are collective operations, such as broadcast and reduce. We have developed MAGPIE, a library of collective communication operations optimized for wide area systems. MAGPIEs algorithms send the minimal amount of data over the slow wide area links, and only incur a single wide area latency. Using our system, existing MPI applications can be run unmodified on geographically distributed systems. On moderate cluster sizes, using a wide area latency of 10 milliseconds and a bandwidth of 1 MByte/s, MAGPIE executes operations up to 10 times faster than MPICH, a widely used MPI implementation; application kernels improve by up to a factor of 4. Due to the structure of our algorithms, MAGPIEs advantage increases for higher wide area latencies.

An efficient implementation of Java's remote method invocation

Java offers interesting opportunities for parallel computing. In particular, Java Remote Method Invocation provides an unusually flexible kind of Remote Procedure Call. Unlike RPC, RMI supports polymorphism, which requires the system to be able to download remote classes into a running application. Suns RMI implementation achieves this kind of flexibility by passing around object type information and processing it at run time, which causes a major run time overhead. Using Suns JDK 1.1.4 on a Pentium Pro/Myri.net cluster, for example, the latency for a null RMI (without parameters or a return value) is 1228 μsec, which is about a factor of 40 higher than that of a user-level RPC. In this paper, we study an alternative approach for implementing RMI, based on native compilation. This approach allows for better optimization, eliminates the need for processing of type information at run time, and makes a light weight communication protocol possible. We have built a Java system based on a native compiler, which supports both compile time and run time generation of marshallers. We find that almost all of the run time overhead of RMI can be pushed to compile time. With this approach, the latency of a null RMI is reduced to 34 μsec, while still supporting polymorphic RMIs (and allowing interoperability with other JVMs).

Best-first fixed-depth minimax algorithms

Abstract This article has three main contributions to our understanding of minimax search: First, a new formulation for Stockmans SSS ∗ algorithm, based on Alpha-Beta, is presented. It solves all the perceived drawbacks of SSS ∗ , finally transforming it into a practical algorithm. In effect, we show that SSS ∗ = Alpha-Beta + transposition tables. The crucial step is the realization that transposition tables contain so-called solution trees, structures that are used in best-first search algorithms like SSS ∗ . Having created a practical version, we present performance measurements with tournament game-playing programs for three different minimax games, yielding results that contradict a number of publications. Second, based on the insights gained in our attempts at understanding SSS ∗ , we present a framework that facilitates the construction of several best-first fixed-depth game-tree search algorithms, known and new. The framework is based on depth-first null-window Alpha-Beta search, enhanced with storage to allow for the refining of previous search results. It focuses attention on the essential differences between algorithms. Third, a new instance of this framework is presented. It performs better than algorithms that are currently used in most state-of-the-art game-playing programs. We provide experimental evidence to explain why this new algorithm, MTD( f ), performs better than other fixed-depth minimax algorithms.

Optimizing parallel applications for wide-area clusters

Recent developments in networking technology cause a growing interest in connecting local area clusters of workstations over wide area links, creating multilevel clusters, or metacomputers. Often, latency and bandwidth of local area and wide area networks differ by two orders of magnitude or more. One would expect only very coarse grain applications to achieve good performance. To test this intuition, we analyze the behavior of several existing medium-grain applications on a wide-area multicluster. We find that high performance can be obtained if the programs are optimized to take the multilevel network structure into account. The optimizations reduce intercluster traffic and hide intercluster latency, and substantially improve performance on wide area multiclusters. As a result, the range of metacomputing applications is larger than previously assumed.

New advances in Alpha-Beta searching

Alpha-Beta has been the algorithm of choice for game-tree search for over three decades. Its success is largely attributable to a variety of enhancements to the basic algorithm that can dramatically improve the search efficiency. Although state-ofthe-art game-playing programs build trees that are close in size to the minimal Alpha-Beta search tree, this paper shows that there is still room for improvement. Three new enhancements are presented: best-first Alpha-Beta search, better use of transpositions, and improvingaspiration search under real-time constraints. Measurements show that these improvements can reduce search effort by 35%.

A New Paradigm for Minimax Search

textabstractThis paper introduces a new paradigm for minimax game-tree search algorithms. MT is a memory-enhanced version of Pearls Test procedure. By changing the way MT is called, a number of best-first game-tree search algorithms can be simply and elegantly constructed (including SSS*). Most of the assessments of minimax search algorithms have been based on simulations. However, these simulations generally do not address two of the key ingredients of high performance game-playing programs: iterative deepening and memory usage. This paper presents experimental data from three game-playing programs (checkers, Othello and chess), covering the range from low to high branching factor. The improved move ordering due to iterative deepening and memory usage results in significantly different results from those portrayed in the literature. Whereas some simulations show alpha-beta expanding almost 100% more leaf nodes than other algorithms [Marsland, Reinefeld & Schaeffer, 1987], our results showed variations of less than 20%. One new instance of our framework MTD(f) out-performs our best alpha-beta searcher (aspiration NegaScout) on leaf nodes, total nodes and execution time. To our knowledge, these are the first reported results that compare both depth-first and best-first algorithms given the same amount of memory.

A performance analysis of transposition-table-driven work scheduling in distributed search

This paper introduces a new scheduling algorithm for parallel single-agent search, transposition table driven work scheduling, that places the transposition table at the heart of the parallel work scheduling. The scheme results in less synchronization overhead, less processor idle time, and less redundant search effort. Measurements on a 128-processor parallel machine show that the scheme achieves nearly-optimal performance and scales well. The algorithm performs a factor of 2.0 to 13.7 times better than traditional work-stealing-based schemes.

Improving multivariate Horner schemes with Monte Carlo tree search

Abstract Optimizing the cost of evaluating a polynomial is a classic problem in computer science. For polynomials in one variable, Horner’s method provides a scheme for producing a computationally efficient form. For multivariate polynomials it is possible to generalize Horner’s method, but this leaves freedom in the order of the variables. Traditionally, greedy schemes like most-occurring variable first are used. This simple textbook algorithm has given remarkably efficient results. Finding better algorithms has proved difficult. In trying to improve upon the greedy scheme we have implemented Monte Carlo tree search, a recent search method from the field of artificial intelligence. This results in better Horner schemes and reduces the cost of evaluating polynomials, sometimes by factors up to two.

Dynamic Scripting with Team Coordination in Air Combat Simulation

Traditionally, behavior of Computer Generated Forces CGFs is controlled through scripts. Building such scripts requires time and expertise, and becomes harder as the domain becomes richer and more life-like. These downsides can be reduced by automatically generating behavior for CGFs using machine learning techniques. This paper focuses on Dynamic Scripting DS, a technique tailored to generating agent behavior. DS searches for an optimal combination of rules from a rule base. Under the assumption that intra-team coordination leads to more effective learning, we propose an extension of DS, called DS+C, with explicit coordination. In a comparison with regular DS we find that the addition of team coordination results in earlier convergence to optimal behavior. In addition, we achieved a performance increase of 20% against an unpredictable opponent. With DS+C, behavior for CGFs can be generated that is more effective since the CGFs act on knowledge achieved by coordination and the behavior converges more efficiently than with regular DS.

Unifying single-agent and two-player search

The seminal works of Nilsson and Pearl in the 1970’s and early 1980’s provide a formal basis for splitting the field of heuristic search into two subfields: single- and two-agent search. The subfields are studied in relative isolation from each other; each having its own distinct character. Despite the separation, a close inspection of the research shows that the two areas have actually been converging. This paper argues that the single/two-agent distinction is not the essence of heuristic search anymore. The state space is characterized by a number of key properties that are defined by the application; single- versus two-agent is just one of many. Both subfields have developed many search enhancements; they are shown to be surprisingly similar and general. Given their importance for creating high performance search applications, it is these enhancements that form the essence of our field. Focusing on their generality emphasizes the opportunity for reuse of the enhancements, allows the field of heuristic search to be redefined as a single unified field, and points the way towards a modern theory of search based on the taxonomy proposed here.