In this paper, we propose a new paradigm for network file system design, serverless network file systems. While traditional network file systems rely on a central server machine, a serverless system utilizes workstations cooperating as peers to provide all file system services. Any machine in the system can store, cache, or control any block of data. Our approach uses this location independence, in combination with fast local area networks, to provide better performance and scalability than traditional file systems. Further, because any machine in the system can assume the responsibilities of a failed component, our serverless design also provides high availability via redundant data storage. To demonstrate our approach, we have implemented a prototype serverless network file system called xFS. Preliminary performance measurements suggest that our architecture achieves its goal of scalability. For instance, in a 32-node xFS system with 32 active clients, each client receives nearly as much read or write throughput as it would see if it were the only active client.

We describe a project to capitalize on newly available levels of computational resources in order to understand human cognition. We are building an integrated physical system including vision, sound input and output, and dextrous manipulation, all controlled by a continuously operating large scale parallel MIMD computer. The resulting system will learn to "think " by building on its bodily experiences to accomplish progressively more abstract tasks. Past experience suggests that in attempting to build such an integrated system we will have to fundamentally change the way artificial intelligence, cognitive science, linguistics, and philosophy think about the organization of intelligence. We expect to be able to better reconcile the theories that will be developed with current work in neuroscience.

We present a distributed algorithm for time-sharing parallel workloads that is competitive with coscheduling. Implicit scheduling allows each local scheduler in the system to make independent decisions that dynamically coordinate the scheduling of cooperating processes across processors. Of particular importance is the blocking algorithm which decides the action of a process waiting for a communication or synchronization event to complete. Through simulation of bulk-synchronous parallel applications, we find that a simple two-phase fixed-spin blocking algorithm performs well; a two-phase adaptive algorithm that gathers run-time data on barrier wait-times performs slightly better. Our results hold for a range of machine parameters and parallel program characteristics. These findings are in direct contrast to the literature that states explicit coscheduling is necessary for fine-grained programs. We show that the choice of the local scheduler is crucial, with a priority-based scheduler p...

This work provides a systematic study of the impact of communication performance on parallel applications in a high performance network of workstations. We develop an experimental system in which the communication latency, overhead, and bandwidth can be independently varied to observe the effects on a wide range of applications. Our results indicate that current efforts to improve cluster communication performance to that of tightly integrated parallel machines results in significantly improved application performance. We show that applications demonstrate strong sensitivity to overhead, slowing down by a factor of 60 on 32 processors when overhead is increased from 3 to 103 s. Applications in this study are also sensitive to per-message bandwidth, but are surprisingly tolerant of increased latency and lower per-byte bandwidth. Finally, most applications demonstrate a highly linear dependence to both overhead and per-message bandwidth, indicating that further improvements in communication performance will continue to improve application performance. 1

Phenomenal improvements in the computational performance of multiprocessors have not been matched by comparable gains in I/O system performance. This imbalance has resulted in I/O becoming a significant bottleneck for many scientific applications. One key to overcoming this bottleneck is improving the performance of parallel file systems. The design of a high-performance parallel file system requires a comprehensive understanding of the expected workload. Unfortunately, until recently, no general workload studies of parallel file systems have been conducted. The goal of the CHARISMA project was to remedy this problem by characterizing the behavior of several production workloads, on different machines, at the level of individual reads and writes. The first set of results from the CHARISMA project describe the workloads observed on an Intel iPSC/860 and a Thinking Machines CM-5. This paper is intended to compare and contrast these two workloads for an understanding of their essential similarities and differences, isolating common trends and platform-dependent variances. Using this comparison, we are able to gain more insight into the general principles that should guide parallel file-system design.

Multiprocessors have permitted astounding increases in computational performance, but many cannot meet the intense I/O requirements of some scientific applications. An important component of any solution to this I/O bottleneck is a parallel file system that can provide high-bandwidth access to tremendous amounts of data in parallel to hundreds or thousands of processors. Most successful systems are based on a solid understanding of the expected workload, but thus far there have been no comprehensive workload characterizations of multiprocessor le systems. This paper presents the results of a three week tracing study in which all file-related activity on a massively parallel computer was recorded. Our instrumentation di ers from previous efforts in that it collects information about every I/O request and about the mix of jobs running in a production environment. We also present the results of a trace-driven caching simulation and recommendations for designers of multiprocessor file systems.

Scheduling algorithms that use application and system knowledge have been shown to be more effective at scheduling parallel jobs on a multiprocessor than algorithms that do not. This paper focuses on obtaining such information for use by a scheduler in a network of workstations environment. The log files from three parallel systems are examined to determine both how to categorize parallel jobs for storage in a job database and what job information would be useful to a scheduler. A Historical Profiler is proposed that stores information about programs and users, and manipulates this information to provide schedulers with execution time predictions. Several preemptive and non-preemptive versions of the FCFS, EASY and Least Work First scheduling algorithms are compared to evaluate the utility of the profiler. It is found that both preemption and the use of application execution time predictions obtained from the Historical Profiler lead to improved performance.

. The Cornell Theory Center (CTC) put a 512-node IBM SP2 system into production in early 1995, and extended traces of batch jobs began to be collected in June of that year. An analysis of the workload shows that it has not only grown, but that its characteristics have changed over time. In particular, job duration increased with time, indicative of an expanding production workload. In addition, there was increasing use of parallelism. As the load has increased and larger jobs have become more frequent, the batch management software (IBM's LoadLeveler) has had difficulty in scheduling the requested resources. New policies were established to improve the situation. This paper will profile how the workload has changed over time and give an in-depth look at the maturing workload. It will examine how frequently certain resources are requested and analyze user submittal patterns. It will also describe the policies that were implemented to improve the scheduling situation and their effect on ...

When multiple jobs compete for processing resources on a parallel computer, the operating system kernel's processor allocation policy determines how many and which processors to allocate to each. This dissertation investigates the issues involved in constructing a processor allocation policy for large scale, message-passing parallel computers supporting a scientific workload. First, the issues that affect the performance of scheduling policies for message-passing parallel systems are examined. We argue that reasonable policies must provide nearly equal resource allocation to all runnable jobs and allocate, to a single job, processors that are in close proximity to one another. Second, the concept of efficiency preservation is defined as a characteristic of processor allocation policie...

In recent years advances in computational speed have been the main focus of research and development in high performance computing. In comparison, the improvement in I/O performance has been modest. Faster processing speeds have created a need for faster I/O as well as for storage and retrieval of vast amounts of data. The technology needed to develop these mass storage systems exists today. Robotic storage libraries are vital components of such systems; however, they normally exhibit high latency and long transmission times. In this paper we analyze the performance of robotic storage libraries and study striping as a technique for improving response time. We show that striping, which improves the effective bandwidth, introduces overhead into the usage of the library's resources, and hence conditions under which it is advantageous are highly dependent on the system's workload. This work was partially done while the author was a summer student at the Lawrence Livermore National Labora...