In this paper, we address the challenge of analyzing simulation data on HPC systems by using Apache Spark, which is a Big Data framework. One of the main problems we encountered with using Spark on HPC systems is the ephemeral data explosion, which is brought about by the curse of persistence in the Spark framework. Data persistence is essential in reducing I/O, but it comes at the cost of storage space. We show that in some cases, Spark scratch data can consume an order of magnitude more space than the input data being analyzed, leading to fatal out-of-disk errors. We investigate the real-world application of scaling machine learning algorithms to predict and analyze failures in multi-physics simulations on 76TB of data (over one trillion training examples). This problem is 2--3 orders of magnitude larger than prior work. Based on extensive experiments at scale, we provide several concrete recommendations as state-of-the-practice, and demonstrate a 7x reduction in disk utilization with negligible increases or even decreases in runtime.

In this article, we investigate limitations in the traditional value-based algorithms for a power-constrained HPC system and evaluate their impact on HPC productivity. We expose the trade-off between allocating system-wide power budget uniformly and greedily under different system-wide power constraints in an oversubscribed system. We experimentally demonstrate that, under the tightest power constraint, the mean productivity of the greedy allocation is 38 percent higher than the uniform allocation whereas, under the intermediate power constraint, the uniform allocation has a mean productivity of 6 percent higher than the greedy allocation. We then propose a new algorithm that adapts its behavior to deliver the combined benefits of the two allocation strategies. We design a methodology with online retraining capability to create application-specific power-execution time models for a class of HPC applications. These models are used in predicting the execution time of an application on the available resources at the time of making scheduling decisions in the power-aware algorithms. We evaluate the proposed algorithm using emulation and simulation environments, and show that our adaptive strategy results in improving HPC resource utilization while delivering a mean productivity that is almost the same as the best performing algorithm across various system-wide power constraints.

In the exascale era, HPC systems are expected to operate under different system-wide power-constraints. For such power-constrained systems, improving per-job flops-per-watt may not be sufficient to improve the total HPC productivity as more number of scientific applications with different compute intensities are migrating to the HPC systems. To measure HPC productivity for such applications, we utilize a monotonically decreasing time-dependent value function, called job-value, with each application. A job-value function represents the value of completing a job for an organization. We begin by exploring the trade-off between two commonly used static power allocation strategies (uniform and greedy) in a power-constrained oversubscribed system. We simulate a large-scale system and demonstrate that, at the tightest power constraint, the greedy allocation can lead to 30% higher productivity compared to the uniform allocation whereas, the uniform allocation can gain up to 6% higher productivity at the relaxed power constraint. We then propose a new dynamic power allocation strategy that utilizes power-performance models derived from offline data. We use these models for reallocating power from running jobs to newly arrived jobs to increase overall system utilization and productivity. In our simulation study, we show that compared to static allocation, the dynamic power allocation policy improves node utilization and job completion rates by 20% and 9%, respectively, at the tightest power constraint. Our dynamic approach consistently earns up to 8% higher productivity compared to the best performing static strategy under different power constraints.

Spatial object association, also referred to as crossmatch of spatial datasets, is the problem of identifying and comparing objects in two or more datasets based on their positions in a common spatial coordinate system. In this work, we evaluate two crossmatch algorithms that are used for astronomical sky surveys, on the following database system architecture configurations: (1) Netezza Performance Serverreg, a parallel database system with active disk style processing capabilities, (2) MySQL Cluster, a high-throughput network database system, and (3) a hybrid configuration consisting of a collection of independent database system instances with data replication support. Our evaluation provides insights about how architectural characteristics of these systems affect the performance of the spatial crossmatch algorithms. We conducted our study using real use-case scenarios borrowed from a large-scale astronomy application known as the large synoptic survey telescope (LSST).

This paper investigates data-preservation, a feature of scientific workflow middleware (SWM) useful for supporting data provenance and "smart recomputation." We observe that in order for an SWM supporting data preservation to achieve decent performance, it should execute on top of copy-on-write file systems. Unfortunately, most file systems in-use at scientific computing facilities were designed without copy-on-write semantics. In response, we design, implement and evaluate a middleware-level solution that is based on user-provided hints and parallelization. The solution can be deployed on top of current file systems and is able to scale almost arbitrarily. Our validation is based on real use-cases from astrophysics and experiments on a cluster with 4 file systems

To provide scientists and engineers with the ability to explore and analyze tera-scale size data-sets we are using a twofold approach. First, we model the data with the objective of creating a compressed yet manageable representation. Second, with that compressed representation, we provide the ability to query the resulting approximation in order to obtain approximate yet sufficient answers; a process called ad-hoc querying. This paper is concerned with a wavelet modeling technique that seeks to capture the important physical characteristics of the target scientific data. Our approach is driven by the compression, which is necessary for viable throughput, along with the end user requirements from the discovery process. Our work contrasts existing research which applies wavelets to range querying, change detection, and clustering problems by working directly with the wavelet decomposition of the data. The difference in this procedure is due primarily to the nature of the data and the requirements of the scientists and engineers. Our approach directly uses the wavelet coefficients of the data to compress as well as query. We describe how the wavelet decomposition is used to facilitate data compression and how queries are posed on the resulting compressed model. Results of this process will be shown for several problems of interest.

As scientific data sets grow exponentially in size, the need for scalable algorithms that heuristically partition the data increases. In this paper, we describe the three-step evolution of a hierarchical partitioning algorithm for large-scale spatio-temporal scientific data sets generated by massive simulations. The first version of our algorithm uses a simple top-down partitioning technique, which divides the data by using a four-way bisection of the spatio-temporal space. The shortcomings of this algorithm lead to the second version of our partitioning algorithm, which uses a bottom-up approach. In this version, a partition hierarchy is constructed by systematically agglomerating the underlying Cartesian grid that is placed on the data. Finally, the third version of our algorithm utilizes the intrinsic topology of the data given in the original scientific problem to build the partition hierarchy in a bottom-up fashion. Specifically, the topology is used to heuristically agglomerate the data at each level of the partition hierarchy. Despite the growing complexity in our algorithms, the third version of our algorithm builds partition hierarchies in less time and is able to build trees for larger size data sets as compared to the previous two versions.

With the advent of fast computer systems, scientists are now able to generate terabytes of simulation data. Unfortunately, the sheer size of these data sets has made efficient exploration of them impossible. To aid scientists in gleaning insight from their simulation data, we have developed an ad-hoc query infrastructure. Our system, called AQSim (short for Ad-hoc Queries for Simulation) reduces the data storage requirements and query access times in two stages. First, it creates and stores mathematical and statistical models of the data at multiple resolutions. Second, it evaluates queries on the models of the data instead of on the entire data set. In this paper, we present two simple but effective statistical modeling techniques for simulation data. Our first modeling technique computes the "true" (unbiased) mean of systematic partitions of the data. It makes no assumptions about the distribution of the data and uses a variant of the root mean square error to evaluate a model. Our second statistical modeling technique uses the Andersen-Darling goodness-of-fit method on systematic partitions of the data. This method evaluates a model by how well it passes the normality test on the data. Both of our statistical models effectively answer range queries. At each resolution of the data, we compute the precision of our answer to the user's query by scaling the one-sided Chebyshev Inequalities with the original mesh's topology. We combine precisions at different resolutions by calculating their weighted average. Our experimental evaluations on two scientific simulation data sets illustrate the value of using these statistical modeling techniques on multiple resolutions of large simulation data sets.

World-Wide Web proxy servers that cache documents can potentially reduce three quantities: the number of requests that reach popular servers, the volume of network traffic resulting from document requests, and the latency that an end-user experiences in retrieving a document. This paper examines the first two using the measures of cache hit rate and weighted hit rate (or fraction of client-requested bytes returned by the proxy). A client request for an uncached document may cause the removal of one or more cached documents. Variable document sizes and types allow a rich variety of policies to select a document for removal, in contrast to policies for CPU caches or demand paging, that manage homogeneous objects. We present a taxonomy of removal policies. Through trace-driven simulation, we determine the maximum possible hit rate and weighted hit rate that a cache could ever achieve, and the removal policy that maximizes hit rate and weighted hit rate. The experiments use five traces of 37 to 185 days of client URL requests. Surprisingly, the criteria used by several proxy-server removal policies (LRU, Hyper-G, and a proposal by Pitkow and Recker) are among the worst performing criteria in our simulation; instead, replacing documents based on size maximizes hit rate in each of the studied workloads.

We describe how to investigate collections of trace data representing network delivery of multimedia information with CHITRA95, a tool that allows a user to visualize, query, statistically analyze and test, transform, and model collections of trace data. CHITRA95 is applied to characterize World Wide Web (WWW) traffic from three workloads: students in a classroom of network-connected workstations, graduate students browsing the Web, undergraduates browsing educational and other materials, as well as traffic on a courseware repository server. We explore the inter-access time of files on a server (i.e., recency), the hit rate from a proxy server cache, and the distributions of file sizes and media types requested. The traffic study also yields statistics on the effectiveness of caching to improve transfer rates. In contrast to past WWW traffic studies, we analyze client as well as server traffic; we compare three workloads rather than drawing conclusions from one workload; and we analyze tcpdump logs to calculate the performance improvement in throughput that an end user sees due to caching.