Networks of workstations are a dominant force in the distributed computing arena, due primarily to the excellent price/performance ratio of such systems when compared to traditionally massively parallel architectures. It is therefore critical to develop programming languages and environments that can potentially harness the raw computational power availab le on these systems. In this article, we present JavaNOW (Java on Networks of Workstations), a Java based framework for parallel programming on networks of workstations. It creates a virtual parallel machine similar to the MPI (Message Passing Interface) model, and provides distributed associative shared memory similar to Linda memory model but with a flexible set of primitive operations. JavaNOW provides a simple yet powerful framework for performing computation on networks of workstations. In addition to the Linda memory model, it provides for shared objects, implicit multithreading, implicit synchronization, object dataflow, and collective communications similar to those defined in MPI. JavaNOW is also a component of the Computational Neighborhood [63], a Java-enabled suite of services for desktop computational sharing. The intent of JavaNOW is to present an environment for parallel computing that is both expressive and reliable and ultimately can deliver good to excellent performance. As JavaNOW is a work in progress, this article emphasizes the expressive potential of the JavaNOW environment and presents preliminary performance results only.

The availability of high-speed networks and increasingly powerful commodity microprocessors is making the usage of clusters, or networks, of computers an appealing vehicle for cost effective parallel computing. Clusters, built using Commodity-Off-The-Shelf (COTS) hardware components as well as free, or commonly used, software, are playing a major role in redefining the concept of supercomputing. In this paper we discuss the reasons why COTS-based clusters are becoming popular environments for running supercomputing applications. We describe the current enabling technologies and present four state-of-theart cluster-based projects. Finally, we summarise our findings and draw a number of conclusions relating to the usefulness and likely future of cluster computing. Copyright © 1999 John Wiley & Sons, Ltd. KEY WORDS: commodity components; clusters; message-passing; supercomputing; parallel computing

Clusters of COTS workstations/PCs are commonly used to implement cost-effective high-performance systems. A central coordinator/manager is often the simplest way to implement many of the operations required for managing these distributed systems. These operations include scheduling of parallel tasks, coordination of access to limited resources, as well as high-level coordination of fault tolerance mechanisms and interactions with external devices. A key disadvantage of using a central manager is that it becomes a critical single point of failure. The UCLA Fault-Tolerant Cluster Testbed (FTCT) project is focused on the implementation of fault-tolerant management for clusters. Unlike most other cluster management projects, our approach is based on active replication and voting and focused on tolerating failure modes other than fail silent and minimizing interruptions to management operations. We describe key aspects of the design and implementation of the FTCT and provide preliminary evaluation of the overheads incurred by our management mechanisms.

With the increased interest in network of workstations for parallel and high performance computing it is necessary to reexamine the use of process migration algorithms, to improve the overall utilization of the system, to achieve high performance and to allow flexible use of idle workstations. Currently, almost all programming environments for parallel systems do not use process migration for task assignments. Instead, a static process assignment is used, with sub optimal performance, especially when several users execute multiple processes simultaneously. This paper highlights the advantages of a process migration scheme for better utilizations of the computing resources as well as to gain substantial speedups in the execution of parallel and multi-tasking applications. We executed several CPU and communication bound benchmarks under PVM, a popular programming environment for parallel computing that uses static process assignment. These benchmarks were executed under the MOSIX multicomputer operating system, with and without its preemptive process migration scheme. The results of these benchmarks prove the advantages of using preemptive process migrations. The paper begins with an overview of MOSIX, a multicomputer enhancement of UNIX that supports transparent process migration for load-balancing, and PVM. We then present the performance of the executions of the benchmarks. Our results show that in some cases the improvements in the performance of PVM with the MOSIX process migration can reach tens or even hundreds of percents.

Active networks allow computations to be performed innetwork at routers as messages pass through them. Active networks offer unique opportunities to optimize networkcentric applications in ways that are not possible using conventional networks. Unfortunately, the need to route packets at full network speed means that very little computation can be done per packet, per router. This seriously restricts the range of in-network applications that can be developed. Computationally intensive applications are restricted to executing outside the network and thus many potential innetwork optimizations are precluded. We propose a scalable cluster architecture using software Distributed Shared Memory (DSM) that can be used as an "attached processor" at routers for executing active code. This novel application of DSM enables the construction of aggressive active network protocols by providing significant compute capacity outside the router's critical packet routing path. The use of DSM simplifies the implementation, and extends the capabilities, of the active packet execution engine in ways that a message passing cluster cannot. Further, the characteristics of active processing enable specific optimizations to consistency maintenance. Keywords Active Networks, Distributed Shared Memory, Cluster Computing, Memory Consistency. 1.

Transactions have been valued for their atomicity and recoverability properties that are useful to several systems, ranging from CAD environment to large-scale databases. Unfortunately, adding transaction support to an existing data repository was traditionally thought to be expensive, mostly due to the fact that the performance of transaction-based systems is usually limited by the performance of the magnetic disks that are used to hold the data repository. In this paper we describe how to use the collective main memory in a Network of Workstations (NOW) to improve the performance of transaction-based systems. We describe the design of our system and its implementation in two independent transaction-based systems, namely EXODUS, and RVM. We evaluate the performance of our prototype using several database benchmarks (like OO7 and TPC-A). Our experimental results indicate that our system delivers up to two orders of magnitude performance improvement compared to its predecessors. 1 Intro...

Chip-multiprocessor (CMP) architectures present a challenge for efficient simulation, combining the requirements of a detailed microprocessor simulator with that of a tightly-coupled parallel system. In this paper, a distributed simulator for target CMPs is presented based on the Message Passing Interface (MPI) designed to run on a host cluster of workstations. Microbenchmark-based evaluation is used to narrow the parallelization design space concerning the performance impact of distributed vs. centralized target L2 simulation, blocking vs. nonblocking remote cache accesses, null-message vs. barrier techniques for clock synchronization, and network interconnect selection. The best combination is shown to yield speedups of up to 16 on a 9-node cluster of dual-CPU workstations, partially due to cache effects.

In the last decade, cluster computing has become the most popular high-performance computing architecture. Although numerous technological innovations have been proposed to improve the interconnection of nodes, many clusters still rely on commodity Ethernet hardware to implement message passing within parallel applications. We present Open-MX, an open-source message passing stack over generic Ethernet. It offers the same abilities as the specialized Myrinet Express stack, without requiring dedicated support from the networking hardware. Open-MX works transparently in the most popular MPI implementations through its MX interface compatibility. It also enables interoperability between hosts running the specialized MX stack and generic Ethernet hosts. We detail how Open-MX copes with the inherent limitations of the Ethernet hardware to satisfy the requirements of message passing by applying an innovative copy offload model. Combined with a careful tuning of the fabric and of the MX wire protocol, Open-MX achieves better performance than TCP implementations, especially on 10 gigabit/s hardware.

In this paper, we describe an active networks architecture for dynamically constructing [clusters of] clusters in response to user requests for computational service. Active network nodes (i.e. routers) match requests for computational service to offers of such service using recent access and usage pattern information to construct on-demand, wide-area clusters. By combining appropriate offers of service with knowledge of current (as well as predictions of future) processor load and network conditions, the system dynamically constructs clusters to cost-effectively meet user needs. The active networks approach exploits the fact that network devices are in an ideal position to decide which resources should be combined to build clusters in response to particular requests. This approach offers advantages in terms of anonymity of service (systems are unaware of one anothers' existence), scalability (scheduling is distributed across many network processors), fault tolerance (service is distributed so failures are recoverable and/or localized) and automatic localization (clusters are created as close to the requestor as possible). Keywords Active Networks, Clusters, Computational Grids, High Performance Computing, Resource Discovery and Allocation. 1

A broad and growing range of possibilities is available to designers of a cluster when choosing an interconnection technology. As the price of network hardware in a cluster can vary from almost free to several thousands of dollars