High-performance distributed computing appears to be shifting away from tightly-connected supercomputers to “computational grids ” composed of heterogeneous systems of networks, computers, storage devices, and various other devices that collectively act as a single geographically distributed “virtual ” computer. One of the great challenges for this environment is providing efficient parallel data access to remote distributed datasets. In this paper, we discuss some of the issues associated with parallel I/O and computatational grids and describe the design of a flexible parallel file system that allows the application to control the behavior and functionality of virtually all aspects of the file system. 1

Accurate simulation of large parallel applications can be facilitated with the use of direct execution and parallel discrete event simulation. This paper describes the use of COMPASS, a direct execution-driven, parallel simulator for performance prediction of programs that include both communication and I/O intensive applications. The simulator has been used to predict the performance of such applications on both distributed memory machines like the IBM SP and shared-memory machines like the SGI Origin 2000. The paper illustrates the usefulness of COMPASS as a versatile performance prediction tool. We use both real-world applications and synthetic benchmarks to study application scalability, sensitivity to communication latency, and the interplay between factors like communication pattern and parallel file system caching on application performance. We also show that the simulator is accurate in its predictions and that it is also efficient in its ability to use parallel si...

Because both processor and interprocessor communication hardware is evolving rapidly with only moderate improvements to file system performance in parallel systems, it is becoming increasingly difficult to provide sufficient input/output (I/O) performance to parallel applications. I/O hardware and file system parallelism are the key to bridging this performance gap. Prerequisite to the development of efficient parallel file system is detailed characterization of the I/O demands of parallel applications. In the paper, we present a comparative study of parallel I/O access patterns, commonly found in I/O intensive scientific applications. The Pablo performance analysis tool and its I/O extensions is a valuable resource in capturing and analyzing the I/O access attributes and their interactions with extant parallel I/O systems. This analysis is instrumental in guiding the development of new application programming interfaces (APIs) for parallel file systems and effective file system polici...

Applications that make use of very large scientific datasets have become an increasingly important subset of scientific applications. In these applications, datasets are often multi-dimensional, i.e., data items are associated with points in a multi-dimensional attribute space, and access to data items is described by range queries. The basic processing involves mapping input data items to output data items, and some form of aggregation of all the input data items that project to the each output data item. We have developed an infrastructure, called the Active Data Repository (ADR), that integrates storage, retrieval and processing of multi-dimensional datasets on distributed-memory parallel architectures with multiple disks attached to each node. In this paper we address efficient execution of range queries on distributed memory parallel machines within ADR framework. We present three potential strategies, and evaluate them under different application scenarios and machine co...

As computational power and storage capacity increase, processing and analyzing large volumes of multi-dimensional datasets play an increasingly important role in many domains of scientific research. Scientific applications that make use of very large scientific datasets have several important characteristics: datasets consist of complex data and are usually multi-dimensional; applications usually retrieve a subset of all the data available in the dataset; various applicationspecific operations are performed on the data items retrieved. Such applications can be supported by object-relational database management systems (OR-DBMSs). In addition to providing functionality to define new complex datatypes and user-defined functions, an OR-DBMS for scientific datasets should contain runtime support that will provide optimized storage for very large datasets and an execution environment for user-defined functions involving expensive operations. In this paper we describe an infrastructure, the ...

This paper presents Clusterfile, a parallel file system that provides parallel file access on a cluster of computers. Existing parallel file systems offer little control over matching the I/O access patterns and file data layout. Without this matching the applications may face the following problems: contention at I/O nodes, fragmentation of file data, false sharing, small network messages, high overhead of scattering /gathering the data. Clusterfile addresses some of these inefficiencies. Parallel applications can physically partition a file in arbitrary patterns. They can also set arbitrary views on a file. Views hide the parallel structure of the file and ease the programmer's burden of computing complex access indices. The intersections between views and layouts are computed by a memory redistribution algorithm. Read and write operations are optimized by pre-computing the direct mapping between access patterns and disks. Clusterfile uses the same data representation for file layouts, access patterns, and the mappings between each other.

Although several algorithms have been developed for the Parallel Disk Model (PDM), few have beenimplemented. Consequently, little has been known about the accuracy of thePDMin measuring I/O time and total running time toperform an out-of-core computation. This paper analyzes timing results on multiple-disk platforms fortwo PDM algorithms, out-of-core radix sort and BMMC permutations, to determine the strengths and weaknesses of thePDM. The results indicate the following. First, good PDM algorithms are usually not I/O bound. Second, of the four PDM parameters, one (problem size) is a good indicator of I/O time and running time, one (memory size) is a good indicator of I/O time but not necessarily running time, and the other two (block size and number of disks) do not necessarily indicate either I/O or running time. Third, because PDM algorithms tendnottobeI/Obound, using asynchronous I/O can reduce I/O wait times signi cantly. The software interface to the PDM is part of the ViC * run-time library. The interface is a set of wrappers that are designed to be both e cient and portable across several underlying le systems and target machines. 1

Performance Availability for Networks of Workstations by Remzi H. Arpaci-Dusseau Software systems for large-scale distributed and parallel machines are difficult to build. When run in dynamic, production environments, not only must such systems perform correctly, but they must also operate with high performance. Much of the previous work in distributed computing has addressed the design of large-scale systems that function correctly, in spite of correctness faults of individual components [18, 49, 82, 86]. However, there has been little development of techniques to tolerate performance faults -- unexpected performance fluctuations from the components that comprise the system. Due to this shortcoming, many systems are overly sensitive to performance variations, in that global performance is high if and only if all system components perform exactly as expected. In this dissertation, we address this deficiency by formalizing the concept of performance availability. Our hypothesis is ...

Exploring and analyzing large volumes of data plays an increasingly important role in many domains of scientific research. We have been developing the Active Data Repository (ADR), an infrastructure that integrates storage, retrieval, and processing of large multi-dimensional scientific datasets on distributed memory parallel machines with multiple disks attached to each node. In earlier work, we proposed three strategies for processing range queries within the ADR framework. Our experimental results show that the relative performance of the strategies changes under varying application characteristics and machine configurations. In this work we investigate approaches to guide and automate the selection of the best strategy for a given application and machine configuration. We describe analytical models to predict the relative performance of the strategies when input data elements are uniformly distributed in the attribute space of the output dataset, restricting the output da...

Many scientific applications have large I/O requirements, in terms of both the size of data and the number of files or data sets. Management, storage, efficient access, and analysis of this data present an extremely challenging task. Traditionally, two different solutions are used for this problem: file I/O or databases. File I/O can provide high performance but is tedious to use with large numbers of files and large and complex data sets. Databases can be convenient, flexible, and powerful but do not perform and scale well for parallel supercomputing applications. We have developed a software system, called Scientific Data Manager (SDM), that aims to combine the good features of both file I/O and databases. SDM provides a high-level API to the user and, internally, uses a parallel file system to store real data and a database to store application-related metadata. SDM takes advantage of various I/O optimizations available in MPI-IO, such as collective I/O and noncontiguous requests, in a manner that is transparent to the user. As a result, users can write and retrieve data with the performance of parallel file I/O, without having to bother with the details of actually performing file I/O. In this paper, we describe the design and implementation of SDM. With the help of two parallel application templates, ASTRO3D and an Euler solver, we illustrate how some of the design criteria affect performance. 0-7803-9802-5/2000/$10.00 c 2000 IEEE