Cloud computing systems fundamentally provide access to large pools of data and computational resources through a variety of interfaces similar in spirit to existing grid and HPC resource management and programming systems. These types of systems offer a new programming target for scalable application developers and have gained popularity over the past few years. However, most cloud computing systems in operation today are proprietary, rely upon infrastructure that is invisible to the research community, or are not explicitly designed to be instrumented and modified by systems researchers. In this work, we present Eucalyptus - an open-source software framework for cloud computing that implements what is commonly referred to as infrastructure as a service (IaaS); systems that give users the ability to run and control entire virtual machine instances deployed across a variety physical resources. We outline the basic principles of the Eucalyptus design, detail important operational aspects of the system, and discuss architectural trade-offs that we have made in order to allow EUCALYPTUS to be portable, modular and simple to use on infrastructure commonly found within academic settings. Finally, we provide evidence that EUCALYPTUS enables users familiar with existing grid and HPC systems to explore new cloud computing functionality while maintaining access to existing, familiar application development software and grid middleware.

Today's scientific workflows use distributed heterogeneous resources through diverse grid and cloud interfaces that are often hard to program. In addition, especially for time-sensitive critical applications, predictable quality of service is necessary across these distributed resources. VGrADS' virtual grid execution system (vgES) provides an uniform qualitative resource abstraction over grid and cloud systems. We apply vgES for scheduling a set of deadline sensitive weather forecasting workflows. Specifically, this paper reports on our experiences with (1) virtualized reservations for batchqueue systems, (2) coordinated usage of TeraGrid (batch queue), Amazon EC2 (cloud), our own clusters (batch queue) and Eucalyptus (cloud) resources, and (3) fault tolerance through automated task replication. The combined effect of these techniques was to enable a new workflow planning method to balance performance, reliability and cost considerations. The results point toward improved resource selection and execution management support for a variety of e-Science applications over grids and cloud systems.

In high-performance computing (HPC) settings, in which multiprocessor machines are shared among users with potentially competing resource demands, processors are allocated to user workload using space sharing. Typically, users interact with a given machine by submitting their jobs to a centralized batch scheduler that implements a site-specific policy designed to maximize machine utilization while providing tolerable turn-around times. To these users, the functioning of the batch scheduler and the policies it implements are both critical operating system components since they control how each job is serviced. In practice, while most HPC systems experience good utilization levels, the amount of time experienced by individual jobs waiting to begin execution has been shown to be highly variable and difficult to predict, leading to user confusion and/or frustration.

Providing QoS (quality of service) in batch resources against the uncertainty of resource availability due to the space-sharing nature of scheduling policies is a critical capability required for high-performance computing. This paper introduces a technique called personal cluster which reserves a partition of batch resources on user's demand in a best-effort manner. A personal cluster provides a private cluster dedicated to the user during a user-specified time period by installing a user-level resource manager on the resource partition. This technique not only enables cost-effective resource utilization and efficient task management but also provides the user a uniform interface to heterogeneous resources regardless of local resource management software. A prototype implementation using a PBS batch resource manager and Globus Toolkits based on Web services shows that the overhead of instantiating a personal cluster of medium size is small, which is just about 1 minute for a personal cluster having 32 processors.

We present a framework for making computation offloading decisions in computational grid settings in which schedulers determine when to move parts of a computation to more capable resources to improve performance. Such schedulers must predict when an offloaded computation will outperform one that is local by forecasting the local cost (execution time for computing locally) and remote cost (execution time for computing remotely and transmission time for the input/output of the computation to/from the remote system). Typically, this decision amounts to predicting the bandwidth between the local and remote systems to estimate these costs. Our framework unifies such decision models by formulating the problem as a statistical decision problem that can either be treated "classically" or using a Bayesian approach. Using an implementation of this framework, we evaluate the efficacy of a number of different decision strategies (several of which have been employed by previous systems). Our results indicate that a Bayesian approach employing automatic change-point detection when estimating the prior distribution is the best-performing approach.

No abstract available.

In this article, we investigate the dynamics exhibited by the production Condor pool at the University of Wisconsin with the goal of understanding its distributional properties. Condor is a cycle-harvesting service originally designed to launch and control "guest" user jobs (in batch mode) on idle workstations. Since its inception in 1985, however, it has expanded to include the ability to run in dedicated mode on clusters, to "glide in" to systems that are not strictly dedicated to Condor, and to "flock" jobs from one site to another based on pre-determined service level agreements (SLAs). Thus it has developed from an enterprise-wide desktop system into a full-fledged global computing infrastructure over its lifetime.

No abstract available.

Most space-sharing parallel computers presently operated by high-performance computing centers use batch-queuing systems to manage processor allocation. In many cases, users wishing to use these batch-queued resources have accounts at multiple sites and have the option of choosing at which site or sites to submit a parallel job. In such a situation, the amount of time a user's job will wait in any one batch queue can significantly impact the overall time a user waits from job submission to job completion. In this work, we explore a new method for providing end-users with predictions for the bounds on the queuing delay individual jobs will experience. We evaluate this method using batch scheduler logs for distributed-memory parallel machines that cover a 9-year period at 7 large HPC centers.Our results show that it is possible to predict delay bounds reliably for jobs in different queues, and for jobs requesting different ranges of processor counts. Using this information, scientific application developers can intelligently decide where to submit their parallel codes in order to minimize overall turnaround time.

Large-scale distributed systems offer computational power at unprecedented levels. In the past, HPC users typically had access to relatively few individual supercomputers and, in general, would assign a one-to-one mapping of applications to machines. Modern HPC users have simultaneous access to a large number of individual machines and are beginning to make use of all of them for single-application execution cycles. One method that application developers have devised in order to take advantage of such systems is to organize an entire application execution cycle as a workflow. The scheduling of such workflows has been the topic of a great deal of research in the past few years and, although very sophisticated algorithms have been devised, a very specific aspect of these distributed systems, namely that most supercomputing resources employ batch queue scheduling software, has heretofore been omitted from consideration, presumably because it is difficult to model accurately. In this work, we augment an existing workflow scheduler through the introduction of methods which make accurate predictions of both the performance of the application on specific hardware, and the amount of time individual workflow tasks will spend waiting in batch queues. Our results show that although a workflow scheduler alone may choose correct task placement based on data locality or network connectivity, this benefit is often compromised by the fact that most jobs submitted to current systems must wait in overcommited batch queues for a significant portion of time. However, incorporating the enhancements we describe improves workflow execution time in settings where batch queues impose significant delays on constituent workflow tasks.

Cycle-harvesting systems such as Condor have been developed to make desktop machines in a local area (which are often similar to clusters in hardware configuration) available as a compute platform. To provide a dual-use capability, opportunistic jobs harvesting cycles from the desktop must be checkpointed before the desktop resources are reclaimed by their owners and the job is evacuated. In this paper, we investigate a new system for computing efficient checkpoint schedules in cycle-harvesting environments. Our system records the historical availability from each resource and fits a statistical model to the observations. Because checkpointing must often traverse the network (i.e. the desktop hosts do not provide sufficient persistent storage for checkpoints), we combine this model with predictions of network performance to the storage site to compute a checkpoint schedule. When an application is initiated on a particular resource, the system uses the computed distribution to parameterize a Markov state-transition model for the application's execution, evaluates the expected time and network overhead as a function of the checkpoint interval, and numerically optimizes with respect to time. We report on the performance of and implementation of this system using the Condor cycle-harvesting environment at the University of Wisconsin. We also evaluate the efficiencies we achieve for a variety of network overheads using trace-based simulation. Finally, we validate our simulations against the observed performance with Condor. Our results indicate that while the choice of model distribution has a relatively small but positive effect on time efficiency, it has a substantial impact on network utilization

In this paper we examine the problem of predicting machine availability in desktop and enterprise computing environments. Predicting the duration that a machine will run until it restarts (availability duration) is critically useful to application scheduling and resource characterization in federated systems. We describe one parametric model fitting technique and two nonparametric prediction techniques, comparing their accuracy in predicting the quantiles of empirically observed machine availability distributions. We describe each method analytically and evaluate its precision using a synthetic trace of machine availability constructed from a known distribution. To detail their practical efficacy, we apply them to machine availability traces from three separate desktop and enterprise computing environments, and evaluate each method in terms of the accuracy with which it predicts availability in a trace driven simulation. Our results indicate that availability duration can be predicted with quantifiable confidence bounds and that these bounds can he used as conservative bounds on lifetime predictions. Moreover a nonparametric method based on a binomial approach generates the most accurate estimates.

We present ARWin, a single user 3D augmented reality desktop. We explain our design considerations and system architecture and discuss a variety of applications and interaction techniques designed to take advantage of this new platform.

In this paper, we describe the use of a cluster as a generalized facility for development. A development facility is a system used primarily for testing and development activities while being operated reliably for many users. We are in the midst of a project to build and operate a large-scale development facility. We discuss our motivation for using clusters in this way and compare the model with a classic computing facility. We describe our experiences and findings from the first phase of this project. Many of these observations are relevant to the design of standard clusters and to future development facilities.

Systems administrators of large clusters often need to perform the same administrative task hundreds or thousands of times. Administrators have traditionally performed some time-consuming tasks, such as operating system installation, configuration, and maintenance, manually. By combining network services such as DHCP, TFTP, FTP, HTTP, and NFS with remote hardware control and scripted installation, configuration, and maintenance techniques, cluster administrators can automate these administrative tasks. Scalable cluster administration addresses this challenge: What hardware and software design techniques can cluster builders use to automate cluster administration on very large clusters? We describe the approach used in the Mathematics and Computer Science Division of Argonne National Laboratory on Chiba City I, a 314-node Linux cluster; and we analyze the scalability, flexibility, performance and reliability benefits and limitations from that approach.

Computational grids provide mechanisms for sharing and accessing large and heterogeneous collections of remote resources such as computers, online instruments, storage space, data, and applications. Resources are requested by specifying a set of desired attributes. Resource attributes have various degrees of dynamism, from mostly static attributes, such as operating system version, to highly dynamic ones, such as available network bandwidth or CPU load. Another dimension of dynamism is introduced by variable and highly diverse sharing policies: resources are made available to the grid community based on locally defined and potentially changing policies.

A critical but often ignored component of system performance is the I/O system. Today's applications demand a great deal from underlying storage systems and software, and both high-performance distributed storage and high level interfaces have been developed to fill these needs.In this paper we discuss the I/O performance of a parallel scientific application on a Linux cluster, the FLASH astrophysics code. This application relies on three I/O software components to provide high-performance parallel I/O on Linux clusters: the Parallel Virtual File System, the ROMIO MPI-IO implementation, and the Hierarchical Data Format library. Through instrumentation of both the application and underlying system software code we discover the location of major software bottlenecks. We work around the most inhibiting of these bottlenecks, showing substantial performance improvement. We point out similarities between the inefficiencies found here and those found in message passing systems, indicating that research in the message passing field could be leveraged to solve similar problems in high-level I/O interfaces.