Scale-up machines perform better for jobs with small and median (KB, MB) data sizes, while scale-out machines perform better for jobs with large (GB, TB) data size. Since a workload usually consists of jobs with different data size levels, we propose building a hybrid Hadoop architecture that includes both scale-up and scale-out machines, which however is not trivial. The first challenge is workload data storage. Thousands of small data size jobs in a workload may overload the limited local disks of scale-up machines. Jobs from scale-up and scale-out machines may both request the same set of data, which leads to data transmission between the machines. The second challenge is to automatically schedule jobs to either scale-up or scale-out cluster to achieve the best performance. We conduct a thorough performance measurement of different applications on scale-up and scale-out clusters, configured with Hadoop Distributed File System (HDFS) and a remote file system (i.e., OFS), respectively. We find that using OFS rather than HDFS can solve the data storage challenge. Also, we identify the factors that determine the performance differences on the scale-up and scale-out clusters and their cross points to make the choice. Accordingly, we design and implement the hybrid scale-up/out Hadoop architecture. Our trace-driven experimental results show that our hybrid architecture outperforms both the traditional Hadoop architecture with HDFS and with OFS in terms of job completion time, throughput and job failure rate.

Many high-performance computing (HPC) sites extend their clusters to support Hadoop MapReduce for a variety of applications. However, HPC cluster differs from Hadoop cluster on the configurations of storage resources. In the Hadoop Distributed File System (HDFS), data resides on the compute nodes, while in the HPC cluster, data is stored on separate nodes dedicated to storage. Dedicated storage offloads I/O load from the compute nodes and provides more powerful storage. Local storage provides better locality and avoids contention for shared storage resources. To gain an insight of the two platforms, in this paper, we investigate the performance and resource utilization of different types (i.e., I/O-intensive, data-intensive and CPU-intensive) of applications on the HPC-based Hadoop platforms with local storage and dedicated storage. We find that the I/O-intensive and data-intensive applications with large input file size can benefit more from the dedicated storage, while these applications with small input file size can benefit more from the local storage. CPU-intensive applications with a large number of small-size input files benefit more from the local storage, while these applications with large-size input files benefit approximately equally from the two platforms. We verify our findings by trace-driven experiments on different types of jobs from the Facebook synthesized trace. This work provides guidance on choosing the best platform to optimize the performance of different types of applications and reduce system overhead.

As parallel file systems span larger and larger numbers of nodes in order to provide the performance and scalability necessary for modern cluster applications, the need for fault-tolerance and high data availability file systems has arisen. Modern parallel file systems spanning tens, hundreds, or even thousands of servers will require fault tolerance to avoid job failure and catastrophic data loss due to a single disk failure or server loss. Effective fault tolerance in parallel file systems must provide a high degree of data resiliency, consistency, and scalable performance. In this paper, we describe a data replication technique that meets the resiliency and consistency requirements of parallel file systems and provides scalable performance. We measure the performance of our proposed mechanism by implementing it in a popular

The trend in parallel computing toward clusters running thousands of cooperating processes per application has led to an I/O bottleneck that has only gotten more severe as the CPU density of clusters has increased. Current parallel file systems provide large amounts of aggregate I/O bandwidth; however, they do not achieve the high degrees of metadata scalability required to manage files distributed across hundreds or thousands of storage nodes. In this paper we examine the use of collective communication between the storage servers to improve the scalability of file metadata operations. In particular, we apply server-to-server communication to simplify consistency checking and improve the performance of file creation, file removal, and file stat. Our results indicate that collective communication is an effective scheme for simplifying consistency checks and significantly improving the performance for several real metadata intensive workloads.

The interaction of simultaneously co-allocated jobs can often create contention in the network infrastructure of a dedicated computational grid. This contention can lead to degraded job run-time performance. We present several bandwidth-aware co-allocating meta-schedulers. These schedulers take into account inter-cluster network utilization as a means by which to mitigate this impact. We make use of a bandwidth-centric parallel job communication model that captures the time-varying utilization of shared inter-cluster network resources. By doing so, we are able to evaluate the performance of grid scheduling algorithms that focus not only on node resource allocation, but also on shared inter-cluster network bandwidth.

In this paper, we describe Arches, an object-oriented framework for building domain-specific PSEs. The framework was designed to support a wide range of problem domains and to be extendable in a way that allows it to target very different high-performance computing models. To demonstrate this flexibility we describe two PSEs that have been developed from the same framework yet solve different problems and target very different computing platforms. The Coven PSE supports parallel applications that need the large-scale parallelism that is found in cost-effective Beowulf clusters. In contrast, the RCADE PSE targets reconfigurable computing (FPGA-based) platforms with fine-grain parallelism. RCADE was designed to aid NASA Earth scientists interested in studying satellite instrument data and who are unlikely to be schooled in low-level hardware design.

Summary form only given. We present a bandwidth-centric parallel job communication model that takes into account inter-cluster network utilization as a means by which to capture the interaction and impact of simultaneously co-allocated jobs in a mini-grid. Our model captures the time-varying utilization of shared inter-cluster network resources in the grid. We compare our dynamic model with previous research that utilizes a fixed execution time penalty for co-allocated jobs. We have found that the fixed penalty model is more generous in its prediction of job turnaround time than our dynamic communication model. Additionally, we see that the penalty co-allocated jobs may experience without causing a severe performance degradation decreases as the number of clusters increases.

Abstract With a focus on commodity PC systems, Beowulf clusters traditionally lack the cutting edge network architectures, memory subsystems, and processor technologies found in their more expensive supercomputer counterparts. Many users find that what Beowulf clusters lack in technology, they more than make up for with their significant cost advantage. In this paper, an architectural extension that adds reconfigurable computing to the network interface of Beowulf clusters is proposed. The proposed extension, called an intelligent network interface card (or INIC), enhances both the network and processor capabilities of the cluster, which has a significant impact on the performance of a crucial class of applications. Furthermore, for some applications, the proposed extension partially compensates for weaknesses in the PC memory subsystem. A prototype of the proposed architecture was constructed and analyzed. In addition, two applications, the 2D Fast Fourier Transform (2D‐FFT) and integer sorting, which benefit from the resulting architecture, are discussed and analyzed on the proposed architecture. Specifically, results indicate that the 2D‐FFT is performed 15–50% faster when using the prototype INIC rather than the comparable Gigabit Ethernet. Early simulation results also indicate that integer sorting will receive a 21–32% performance boost from the prototype. The significant improvements seen with a relatively limited prototype lead to the conclusion that cluster network interfaces enhanced with reconfigurable computing could significantly improve the Beowulf architecture. Copyright © 2003 John Wiley & Sons, Ltd.

Parallel I/O remains a critical problem for cluster computing. A significant number of important applications need high performance parallel I/O and most cluster systems provide enough hardware to deliver the required performance. System software for achieving the desired goals remains in the research and development stage. A number of parallel file systems have achieved remarkable goals in one or more of several key areas related to parallel I/O, but there is still great reluctance to commit to any file system currently available. This is mostly due to the fact that these file systems do not address enough issues at once in a package that is robust enough for widespread use. Critical goals in the development of an operation parallel file system for clusters include: high performance with scalability; reliability/fault tolerance; flexible and efficient integration with parallel codes; portability. These issues give rise to problems with interfaces and semantics, in addition to specific technical problems such as distributed locking, caching, and redundancy. The next generation of parallel file systems must look beyond traditional interfaces, semantics, and implementation methods in order achieve the desired goals. Of equal importance is the issue of knowing to what extent a given file system achieves these goals. Given that no file system is likely to address all of these goals equally well, it is important to be able to measure a given file system's utility in these areas through benchmarking or other evaluation methods. We explore a few of these issues and include specific examples and a case study of the PVFS V2 team's approach to these issues.

Presents the design of the Coven framework for construction of problem solving environments (PSEs) for parallel computers. PSEs are an integral part of modern high performance computing (HPC) and Coven attempts to simplify PSE construction. Coven targets Beowulf cluster parallel computers but independent of any particular domain for the PSE. Multithreaded parallel applications are created with Coven that are capable of supporting most of the constructs in a typical parallel programming language. Coven uses an agent-based front-end which allows multiple custom interfaces to be constructed. Examples of the use of Coven in the construction of prototype PSEs are shown, and the effectiveness of these PSEs is evaluated in terms of the performance of the applications they generate.

With a focus on commodity PC systems, Beowulf clusters traditionally lack the cutting edge network architectures, memory subsystems, and processor technologies found in their more expensive supercomputer counterparts. What Beowulf clusters lack in technology, they more than make up for with their significant cost advantage over traditional supercomputers. We propose an architectural extension that adds reconfigurable computing to the network interface of Beowulf clusters. This enhances both the network and processor capabilities of the cluster. Furthermore, for some applications, the proposed extension partially compensates for weaknesses in the PC memory subsystem. We discuss two applications, the 2D Fast Fourier Transform (FFT) and integer sorting, which benefit from the resulting architecture.

With a focus on commodity PC systems, Beowulf clusters traditionally lack the cutting edge network architectures, memory subsystems, and processor technologies found in their more expensive supercomputer counterparts. What Beowulf clusters lack in technology, they more than make up for with their significant cost advantage over traditional supercomputers. This paper presents the cost implications of an architectural extension that adds reconfigurable computing to the network interface of Beowulf clusters. A quantitative idea of cost-effectiveness is formulated to evaluate computing technologies. Here, cost-effectiveness is considered in the context of two applications: the 2D Fast Fourier Transform (2D-FFT) and integer sorting.

Integral equations solved by the method of moments (MOM) are an important and computationally intense class of problems in antenna design and other areas of electromagnetics. Particularly when structures become electrically large, MOM solutions become intractable as they lead to large, densely-filled, complex matrices, the solution of which is numerically and computationally intensive. Several numerical techniques have been applied to make these solutions more tractable, notably spatial decomposition and a multigrain method known as reduced current fidelity. This paper investigates the impact of these techniques on parallel computing approaches to solving this class of problem, using networks of workstations (NOWs). The resulting effects on the efficiency and communication patterns of the parallel programs are explored.

The use of reconfigurable hardware to perform high precision operations such as IEEE floating point operations has been limited in the past by FPGA resources. We discuss the implementation of IEEE single precision floating-point multiplication and addition. Then, we assess the practical implications of using these operations in the Xilinx 4000 series FPGAs considering densities available now and scheduled for the near future. For each operation, we present space requirements and performance information. This is followed by a discussion of an algorithm, matrix multiplication, based on these operations, which achieves performance comparable to conventional microprocessors. Algorithm implementation options and their performance implications are discussed and corresponding measured results are given.

lnterconnection network design plays a central role in the design of parallel systems. Most of the previous research has evaluated the performance of interconnection networks in isolation. In this study, we investigate the relationship between application program characteristics and interconnection network performance using an execution driven simulation test bed: the Reconfigurable Architecture Workbench (RAW). We simulate five topological configurations of a k-ary n-cube interconnect and four different network link models for a 4,096 node SIMD machine, and quantify the impact of the network on two application programs. We provide experimental evidence that such "in-context" simulation provides a better view of the impact of network design variables on system performance. We show that recent results, indicating that low-dimensional designs provide better ICN performance, ignore application requirements that may favor high-dimensional designs. Furthermore, applications that would appear to favor low dimensional designs may not, in fact, be significantly impacted by the network's dimensionality. We experimentally test the results of published performance models comparing the use of a synthetic load to that of a load generated by a typical application program.

Dedicated cluster parallel computers (DCPCs) are emerging as low-cost high performance environments for many important applications in science and engineering. A significant class of applications that perform well on a DCPC are coarse-grain applications that involve large amounts of file I/O. Current research in parallel file systems for distributed systems is providing a mechanism for adapting these applications to the DCPC environment. We present the Parallel Virtual File System (PVFS), a system that provides disk striping across multiple nodes in a distributed parallel computer and file partitioning among tasks in a parallel program. PVFS is unique among similar systems in that it uses a stream-based approach that represents each file access with a single set of request parameters and decouples the number of network messages from details of the file striping and partitioning. PVFS also provides support for efficient collective file accesses and allows overlapping file partitions. We present results of early performance experiments that show PVFS achieves excellent speedups in accessing moderately sized file segments.