arXiv subjects

Magdalena Balazinska

Publications and source records attributed to Magdalena Balazinska.

27 records · Page 2Linked to original sources

arXiv2016

Comparative Evaluation of Big-Data Systems on Scientific Image Analytics Workloads

Scientific discoveries are increasingly driven by analyzing large volumes of image data. Many new libraries and specialized database management systems (DBMSs) have emerged to support such tasks. It is unclear, however, how well these systems support real-world image analysis use cases, and how performant are the image analytics tasks implemented on top of such systems. In this paper, we present the first comprehensive evaluation of large-scale image analysis systems using two real-world scientific image data processing use cases. We evaluate five representative systems (SciDB, Myria, Spark, Dask, and TensorFlow) and find that each of them has shortcomings that complicate implementation or hurt performance. Such shortcomings lead to new research opportunities in making large-scale image analysis both efficient and easy to use.

cs.DB↗

arXiv2016

View-Driven Deduplication with Active Learning

Visual analytics systems such as Tableau are increasingly popular for interactive data exploration. These tools, however, do not currently assist users with detecting or resolving potential data quality problems including the well-known deduplication problem. Recent approaches for deduplication focus on cleaning entire datasets and commonly require hundreds to thousands of user labels. In this paper, we address the problem of deduplication in the context of visual data analytics. We present a new approach for record deduplication that strives to produce the cleanest view possible with a limited budget for data labeling. The key idea behind our approach is to consider the impact that individual tuples have on a visualization and to monitor how the view changes during cleaning. With experiments on nine different visualizations for two real-world datasets, we show that our approach produces significantly cleaner views for small labeling budgets than state-of-the-art alternatives and that it also stops the cleaning process after requesting fewer labels.

cs.DB↗

arXiv2016

PerfEnforce: A Dynamic Scaling Engine for Analytics with Performance Guarantees

In this paper, we present PerfEnforce, a scaling engine designed to enable cloud providers to sell performance levels for data analytics cloud services. PerfEnforce scales a cluster of virtual machines allocated to a user in a way that minimizes cost while probabilistically meeting the query runtime guarantees offered by a service level agreement. With PerfEnforce, we show how to scale a cluster in a way that minimally disrupts a user's query session. We further show when to scale the cluster using one of three methods: feedback control, reinforcement learning, or perceptron learning. We find that perceptron learning outperforms the other two methods when making cluster scaling decisions.

cs.DB↗

arXiv2016

PipeGen: Data Pipe Generator for Hybrid Analytics

We develop a tool called PipeGen for efficient data transfer between database management systems (DBMSs). PipeGen targets data analytics workloads on shared-nothing engines. It supports scenarios where users seek to perform different parts of an analysis in different DBMSs or want to combine and analyze data stored in different systems. The systems may be colocated in the same cluster or may be in different clusters. To achieve high performance, PipeGen leverages the ability of all DBMSs to export, possibly in parallel, data into a common data format, such as CSV or JSON. It automatically extends these import and export functions with efficient binary data transfer capabilities that avoid materializing the transmitted data on the file system. We implement a prototype of PipeGen and evaluate it by automatically generating data pipes between five different DBMSs. Our experiments show that PipeGen delivers speedups up to 3.8x compared with manually exporting and importing data across systems using CSV.

cs.DB↗

arXiv2015

Efficient Iterative Processing in the SciDB Parallel Array Engine

Many scientific data-intensive applications perform iterative computations on array data. There exist multiple engines specialized for array processing. These engines efficiently support various types of operations, but none includes native support for iterative processing. In this paper, we develop a model for iterative array computations and a series of optimizations. We evaluate the benefits of an optimized, native support for iterative array processing on the SciDB engine and real workloads from the astronomy domain.

cs.DB↗

arXiv2012

PerfXplain: Debugging MapReduce Job Performance

While users today have access to many tools that assist in performing large scale data analysis tasks, understanding the performance characteristics of their parallel computations, such as MapReduce jobs, remains difficult. We present PerfXplain, a system that enables users to ask questions about the relative performances (i.e., runtimes) of pairs of MapReduce jobs. PerfXplain provides a new query language for articulating performance queries and an algorithm for generating explanations from a log of past MapReduce job executions. We formally define the notion of an explanation together with three metrics, relevance, precision, and generality, that measure explanation quality. We present the explanation-generation algorithm based on techniques related to decision-tree building. We evaluate the approach on a log of past executions on Amazon EC2, and show that our approach can generate quality explanations, outperforming two naive explanation-generation methods.

cs.DB↗

arXiv2012

How to Price Shared Optimizations in the Cloud

Data-management-as-a-service systems are increasingly being used in collaborative settings, where multiple users access common datasets. Cloud providers have the choice to implement various optimizations, such as indexing or materialized views, to accelerate queries over these datasets. Each optimization carries a cost and may benefit multiple users. This creates a major challenge: how to select which optimizations to perform and how to share their cost among users. The problem is especially challenging when users are selfish and will only report their true values for different optimizations if doing so maximizes their utility. In this paper, we present a new approach for selecting and pricing shared optimizations by using Mechanism Design. We first show how to apply the Shapley Value Mechanism to the simple case of selecting and pricing additive optimizations, assuming an offline game where all users access the service for the same time-period. Second, we extend the approach to online scenarios where users come and go. Finally, we consider the case of substitutive optimizations. We show analytically that our mechanisms induce truth- fulness and recover the optimization costs. We also show experimentally that our mechanisms yield higher utility than the state-of-the-art approach based on regret accumulation.

cs.DB↗

arXiv2010

Astronomy in the Cloud: Using MapReduce for Image Coaddition

In the coming decade, astronomical surveys of the sky will generate tens of terabytes of images and detect hundreds of millions of sources every night. The study of these sources will involve computation challenges such as anomaly detection and classification, and moving object tracking. Since such studies benefit from the highest quality data, methods such as image coaddition (stacking) will be a critical preprocessing step prior to scientific investigation. With a requirement that these images be analyzed on a nightly basis to identify moving sources or transient objects, these data streams present many computational challenges. Given the quantity of data involved, the computational load of these problems can only be addressed by distributing the workload over a large number of nodes. However, the high data throughput demanded by these applications may present scalability challenges for certain storage architectures. One scalable data-processing method that has emerged in recent years is MapReduce, and in this paper we focus on its popular open-source implementation called Hadoop. In the Hadoop framework, the data is partitioned among storage attached directly to worker nodes, and the processing workload is scheduled in parallel on the nodes that contain the required input data. A further motivation for using Hadoop is that it allows us to exploit cloud computing resources, e.g., Amazon's EC2. We report on our experience implementing a scalable image-processing pipeline for the SDSS imaging database using Hadoop. This multi-terabyte imaging dataset provides a good testbed for algorithm development since its scope and structure approximate future surveys. First, we describe MapReduce and how we adapted image coaddition to the MapReduce framework. Then we describe a number of optimizations to our basic approach and report experimental results comparing their performance.

cs.DC↗

arXiv2009

Believe It or Not: Adding Belief Annotations to Databases

We propose a database model that allows users to annotate data with belief statements. Our motivation comes from scientific database applications where a community of users is working together to assemble, revise, and curate a shared data repository. As the community accumulates knowledge and the database content evolves over time, it may contain conflicting information and members can disagree on the information it should store. For example, Alice may believe that a tuple should be in the database, whereas Bob disagrees. He may also insert the reason why he thinks Alice believes the tuple should be in the database, and explain what he thinks the correct tuple should be instead. We propose a formal model for Belief Databases that interprets users' annotations as belief statements. These annotations can refer both to the base data and to other annotations. We give a formal semantics based on a fragment of multi-agent epistemic logic and define a query language over belief databases. We then prove a key technical result, stating that every belief database can be encoded as a canonical Kripke structure. We use this structure to describe a relational representation of belief databases, and give an algorithm for translating queries over the belief database into standard relational queries. Finally, we report early experimental results with our prototype implementation on synthetic data.

cs.DB↗