Cluster Of Workstations Research Articles

Trading Replication for Communication in Parallel Distributed-Memory Dense Solvers

We present new communication-efficient parallel dense linear solvers: a solver for triangular linear systems with multiple right-hand sides and an LU factorization algorithm. These solvers are highly parallel and they perform a factor of 0.4P1/6 less communication than existing algorithms, where P is number of processors. The new solvers reduce communication at the expense of using more temporary storage. Previously, algorithms that reduce communication by using more memory were only known for matrix multiplication. Our algorithms are recursive, elegant, and relatively simple to implement. We have implemented them using MPI, a message-passing libray, and tested them on a cluster of workstations.

DISTRIBUTED PARALLEL PROCESSING TECHNIQUES FOR ADAPTIVE SONAR BEAMFORMING

Quiet submarine threats and high clutter in the littoral environment increase computation and communication demands on beamforming arrays, particularly for applications that require in-array autonomous operation. By coupling each transducer node in a distributed array with a microprocessor, and networking them together, embedded parallel processing for adaptive beamformers can glean advantages in execution speed, fault tolerance, scalability, power, and cost. In this paper, a novel set of techniques for the parallelization of adaptive beamforming algorithms is introduced for in-array sonar signal processing. A narrowband, unconstrained, Minimum Variance Distortionless Response (MVDR) beamformer is used as a baseline to investigate the efficiency and effectiveness of this method in an experimental fashion. Performance results are also included, among them execution times, parallel efficiencies, and memory requirements, using a distributed system testbed comprised of a cluster of workstations connected by a conventional network.

Dynamic load balancing for parallel interval-Newton using message passing

Branch-and-prune and branch-and-bound techniques are commonly used for intelligent search in finding all solutions, or the optimal solution, within a space of interest. The corresponding binary tree structure provides a natural parallelism allowing concurrent evaluation of subproblems using parallel computing technology. Of special interest here are techniques derived from interval analysis, in particular an interval-Newton/generalized-bisection procedure. In this context, we discuss issues of load balancing and work scheduling that arise in the implementation of parallel interval-Newton on a cluster of workstations using message passing, and describe and analyze techniques for this purpose. Results using an asynchronous diffusive load balancing strategy show that a consistently high efficiency can be achieved in solving nonlinear equations, providing excellent scalability, especially with the use of a two-dimensional torus virtual network. The effectiveness of the approach used, especially in connection with a novel stack management scheme, is also demonstrated in the consistent superlinear speedups observed in performing global optimization.

Date movement and control substrate for parallel adaptive applications

AbstractIn this paper, we present the Data Movement and Control Substrate (DMCS), a library which implements low‐latency one‐sided communication primitives for use in parallel adaptive and irregular applications. DMCS is built on top of low‐level, vendor‐specific communication subsystems such as LAPI (Low‐level Application Programme Interface) for IBM SP machines, as well as on widely available message‐passing libraries like MPI for clusters of workstations and PCs. DMCS adds a small overhead to the communication operations provided by the lower communication system. In return, DMCS provides a flexible and easy to understand application program interface for one‐sided communication operations. Furthermore, DMCS is designed so that it can be easily ported and maintained by non‐experts. Copyright © 2002 John Wiley & Sons, Ltd.

Solving Systems of Linear Equations with Relaxed Monte Carlo Method

The problem of solving systems of linear algebraic equations by parallel Monte Carlo numerical methods is considered. A parallel Monte Carlo method with relaxation is presented. This is a report of a research in progress, showing the effectiveness of this algorithm. Theoretical justification of this algorithm and numerical experiments are presented. The algorithms were implemented on a cluster of workstations using MPI.

An Adaptive Parallel Distributive Join Algorithm on a Cluster of Workstations

In this paper, we present an adaptive version of the parallel Distributive Join (DJ) algorithm that we proposed in l5r. The adaptive parallel DJ algorithm can handle the data skew in operand relations efficiently. We implemented the original and adaptive parallel DJ algorithms on a network of Alpha workstations using the Parallel Virtual Machine (PVM). We analyzed the performance of the algorithms, and compared it with that of the parallel Hybrid-Hash (HH) join algorithms. Our results show that the parallel DJ algorithms perform comparably with the parallel HH join algorithms over the entire range of the number of processors used and for different join selectivities. A significant advantage of the parallel DJ algorithms is that they can easily support non-equijoin operations.

High-level Language Support for User-defined Reductions

The optimized handling of reductions on parallel supercomputers or clusters of workstations is critical to high performance because reductions are common in scientific codes and a potential source ...

PHOENIX: A Self Adaptable Monitoring Platform for Cluster Management

Distributed systems based on cluster of workstation are more and more difficult to manage due to the increasing number of processors involved, and the complexity of associated applications. Such systems need efficient and flexible monitoring mechanisms to fulfill administration services requirements. In this paper, we present PHOENIX a distributed platform supporting both applications and operating system monitoring with a variable granularity. The granularity is defined using logical expressions to specify complex monitoring conditions. These conditions can be dynamically modified during the application execution. Observation techniques, based on an automatic probe insertion combined with a system agent to minimize the PHOENIX execution time overhead. The platform extensibility offers a suitable environment to design distributed value added services (performance monitoring, load balancing, accounting, cluster management, etc.).

Fair Scheduling of General-Purpose Workloads on Workstation Clusters

In this paper we present a scheduling strategy for workstation clusters able to effectively and fairly schedule general-purpose workloads potentially made up by compute-bound, interactive, and I/O-intensive applications, that may each be sequential, client-server, or parallel. The scheduling strategy allocates resources to processes of the same parallel applications in such a way that they all get the same CPU share regardless of the level of resource contention on the respective machines, and relies on an extended i>stride scheduler to fairly allocate individual workstations. A simulation analysis carried out for a variety of workloads and operational conditions shows that our strategy (a) delivers good performance to all the applications classes composing general-purpose workloads, (b) fairly allocates resources among competing applications, and (c) outperforms alternative strategies.

Broadcast Scheduling Optimization for Heterogeneous Cluster Systems

Network of workstation (NOW) is a cost-effective alternative to massively parallel supercomputers. As commercially available off-the-shelf processors become cheaper and faster, it is now possible to build a PC or workstation cluster that provides high computing power within a limited budget. However, a cluster may consist of different types of processors, and this heterogeneity within a cluster complicates the design of efficient collective communication protocols. This paper shows that a simple heuristic called fastest-node-first (FNF) (1998, M. Banikazemi, V. Moorthy, and D. K. Panda, in “Proceedings of the International Parallel Processing Conference”) is very effective in reducing the broadcast time for heterogeneous cluster systems. Despite the fact that the FNF heuristic fails to give the optimal broadcast time for a general heterogeneous network of workstations, we prove that FNF always gives the optimal broadcast time in several special cases of clusters. Based on these special case results, we show that FNF is an approximation algorithm that guarantees a competitive ratio of 2. From these theoretical results we also derive techniques to speed up the branch-and-bound search for the optimal broadcast schedule in HNOW.

Parallelizing Tabu Search on a Cluster of Heterogeneous Workstations

In this paper, we present the parallelization of tabu search on a network of workstations using PVM. Two parallelization strategies are integrated: functional decomposition strategy and multi-search threads strategy. In addition, domain decomposition strategy is implemented probabilistically. The performance of each strategy is observed and analyzed. The goal of parallelization is to speedup the search in finding better quality solutions. Observations support that both parallelization strategies are beneficial, with functional decomposition producing slightly better results. Experiments were conducted for the VLSI cell placement, an NP-hard problem, and the objective was to achieve the best possible solution in terms of interconnection length, timing performance (circuit speed), and area. The multiobjective nature of this problem is addressed using a fuzzy goal-based cost computation.

On the Performance and Scalability of NFS over a Myrinet-based Network of Workstations

The development of high performance networks and multimedia applications has spurred the need of developing more efficient distributed file systems. In the past, distributed file systems, such as NFS (Network File System), have been developed aiming to provide transparent remote access to shared file systems across computer platforms. Even though, NFS has been widely studied and new releases are being developed, there are still some open issues regarding its performance and scalability. In this work we explore the operation and performance of NFS when used to support parallel data intensive-processing applications over a cluster of workstations interconnected by a 1-Gbps network. We first consider a parallel implementation of a video compression software encoder. The application is characterized by the large amount of data to be distributed to the processing workstations. Through an exhaustive campaign of experiments, we evaluate the impact of several system key parameters over the performance and scalability of NFS. We then explore the operation of NFS using a synthetic load pattern. Our main results provide the guidelines to set-up some of the key parameters for a more efficient operation of NFS over a cluster of workstations interconnected by a high-speed interconnection network.

Parallel numerical modelling of the Antarctic Ice Sheet

The Antarctic Ice Sheet comprises the West Antarctic Ice Sheet and the much larger East Antarctic Ice Sheet. Fast flowing ice streams and outlet glaciers are important dynamic components of the ice sheet system, and a grid resolution of at least 20 km is required to identify many of these areas. Previous fine resolution numerical models have focussed on ice flow in West Antarctica or on fine resolution modelling of subsections of the ice sheet, since the size of East Antarctica has generally precluded studies of the whole ice sheet at a resolution adequate to identify complex flow features. The equations describing ice flow are highly non-linear, making this a computationally intensive problem. We use a staggered grid for calculation of ice diffusivity to overcome numerical instability, and a sparse packing scheme to take account of the irregular boundary of Antarctica. We have developed an efficient parallel temperature-dependent ice flow model of the entire grounded portion of the Antarctic Ice Sheet at a resolution of 20 km. The model was primarily written to run on a commodity cluster of workstations, and performance results for this and other systems are presented. Ice flow patterns at steady state compare well with recently published balance velocity calculations.

Computation of two-phase flow in steam generator using Domain Decomposition and Local Zoom methods

We present flow simulations in the Steam Generator of a pressurized water nuclear reactor using Domain Decomposition Methods (DDM) and Local Zoom methods on workstation cluster. Concerning the DDM, we use a Dirichlet–Neumann approach jointly with FEM for averaged mixture balance equations. The algorithm, based on parallel or sequential iteration-by-subdomain method, works with overlapping or nonoverlapping subdomains and with conforming or nonconforming meshing. With DDM, the computational problem size is easily increased to about 100,000 mesh cells and the CPU time is strongly reduced. Concerning the Local Zoom computations, the used Local Defect Correction (LDC) method, in 3D local hierarchical multigrid context, is shown. The LDC computation results are compared with the classical full domain computation results (with high or low spatial resolution). We conclude in an improvement of the accuracy on the full domain with a high coherence between the zoom and the full domain.

MPI–Delphi: an MPI implementation for visual programming environments and heterogeneous computing

The goal of a parallel program is to reduce the execution time, compared to the fastest sequential program solving the same problem. Parallel programming is growing due to the widespread use of network of workstations (NOWs) or powerful PCs in high-performance computing. Because the hardware components are all commodity devices, NOWs are much more cost-effective than custom machines with similar technology. In this environment, the typical programming model used has been message-passing and the MPI library has become the standard in the distributed-memory computing model. On the other hand, visual programming environments try to simply the task of developing applications. They provide programmers with several standard components for creating programs. Delphi constitutes one of the most popular visual programming environments nowadays in the Windows market place. In this paper, we present MPI–Delphi, an implementation of MPI for writing parallel applications using Delphi visual programming environment. We show how MPI–Delphi has been developed, and how it makes possible to manage a cluster of homogeneous/heterogeneous PCs. Two examples of use of MPI–Delphi in a heterogeneous cluster of workstations with a mixture of Windows and Linux operating systems are also included. The MPI–Delphi interface is suitable for some specific kinds of problems, such as monitoring parallel programs of long execution time, or computationally intensive graphical simulations. In addition, MPI–Delphi has proven to be a good tool for research, as the development of new algorithms can be carried out quickly and, therefore, time spent on the debugging of such algorithms is reduced. Finally, we conclude by explaining some of the tasks we think MPI–Delphi is suitable for.

SIMULATION OF HIERARCHICAL RESOURCE MANAGEMENT FOR META-COMPUTING SYSTEMS

Optimal scheduling in meta-computing environments still is an open research question. Various resource management (RM) architectures have been proposed in the literature (e.g. [2][13][12]). In the present paper we explore, through simulation, various multi-level scheduling strategies for compound computing environments comprising several clusters of workstations. We study global and local RM and their interaction. The local RM comprises both the cluster management and operating system schedulers. Each level refines the scheduling decisions of the layer above it, taking into account the latest resource information. Our experiments explore conventional strategies like First Come, First Served (FCFS) and Shortest Job First (SJF) at the global RM level. At all levels, the schedulers strive to maintain a good load balance. The unit of load balancing at the global level is the job consisting of one or more parallel tasks; at the local level it is the task. The results of our simulations indicate that, especially at high system loads, the use of a global RM can result in a significant performance gain.

A new fast message passing communication system for multiprocessor workstation clusters

In this paper, a fast message passing communication architecture (FMP) is proposed in order to reduce the software overhead of communication, which has been a great impediment to the good performance of workstation clusters. A FMP system is implemented over the SUN Ultra2-Myrinet platform. Measurements show that this implementation has achieved a one-way latency of 11.2us for one-byte packets in network communication, and only 4.9us in local communication. Its bandwidth can reach as high as 338Mb/s for 8KB packets in network communication, and more than 770Mb/s in local message passing. These results tell us that FMP can really exploit the performance of supercomputers and high-speed networks. Local communication is implemented through shared memory within a single host, while the methods to decrease overheads of network communication depend on user-level communication protocol, pipeline transmission, credit flow control and multithreading. Furthermore, the whole system provides the interface for both network and local communication.

Sound from Turbulence Convected by a Parallel Flow Within a Rectangular Duct

A previously derived high-frequency solution to Lilley's equation is used to study the far-field characteristics of the sound generated by a turbulent flow within a semi-infinite duct of nonaxisymmetric cross section. The source term is simplified by assuming the turbulence to be axisymmetric about the mean flow direction, and a relatively simple formula for the far-field azimuthal and polar sound field variation is obtained. The numerical computations involve the calculation of ray trajectories, and a parallel computer code has been developed for this purpose. It is expected that the code will be able to produce azimuthal directivity patterns in a few minutes on a large parallel computer system. The method has been applied to a rectangular duct, and numerical results for the azimuthal and polar variation of the sound field were computed on a cluster of workstations for a number of duct aspect ratios, wall impedances values, and source locations

Targeting Ada95/DSA for distributed simulation of multiprotocol communication networks

This paper describes the use of Ada95/DSA as middleware for modeling network's entities into distributed objects which can be used for both distributed execution and simulation without transformation of the model. First, general issues of simulation of network's entities are briefly presented. Then, the paper shows how DSA features are used to abstract these entities in a manner as they can be executed or simulated on a cluster of workstations. The DSA based model of distributed simulation is built on a combination of two kinds of classes: the first one support general purpose of distributed objects computing and the second deal with the development of distributed network elements. To illustrate the feasibility of this approach, a partial implementation of classes that have been used to model routers running a subset of border gateway protocol (BGP) is presented.

Wrapping MPI-based legacy codes as Java/CORBA components

Techniques for wrapping an MPI-based molecular dynamics (MD) simulation code as Java/CORBA components, for use within a distributed component based problem solving environment (CB-PSE), is presented. A legacy code for simulating a Lennard–Jones fluid is first wrapped as a single CORBA object, followed by division of the code into computational sub-units, where each sub-unit is wrapped as a CORBA object containing MPI calls, and run on a cluster of workstations — enabling different MPI implementations to inter-operate. Using a Java implementation, users can submit simulation tasks through a Web based interface, without needing to know implementation details of the legacy code, or the exact interaction between sub-units within the code. We provide performance comparisons of wrapping the entire MD code as a single object versus wrapping sub-units within it, and offer a simple performance model to explain our findings.