Zacharia Fadika

DELMA: Dynamically ELastic MapReduce Framework for CPU-Intensive Applications

Since its introduction, MapReduce implementations have been primarily focused towards static compute cluster sizes. In this paper, we introduce the concept of dynamic elasticity to MapReduce. We present the design decisions and implementation tradeoffs for DELMA, (Dynamically Elastic MapReduce), a framework that follows the MapReduce paradigm, just like Hadoop MapReduce, but that is capable of growing and shrinking its cluster size, as jobs are underway. In our study, we test DELMA in diverse performance scenarios, ranging from diverse node additions to node additions at various points in the application run-time with various dataset sizes. The applicability of the MapReduce paradigm extends far beyond its use with large-scale data intensive applications, and can also be brought to bear in processing long running distributed applications executing on small-sized clusters. In this work, we focus both on the performance of processing hierarchical data in distributed scientific applications, as well as the processing of smaller but demanding input sizes primarily used in small clusters. We run experiments for datasets that require CPU intensive processing, ranging in size from Millions of input data elements to process, up to over half a billion elements, and observe the positive scalability patterns exhibited by the system. We show that for such sizes, performance increases accordingly with data and cluster size increases. We conclude on the benefits of providing MapReduce with the capability of dynamically growing and shrinking its cluster configuration by adding and removing nodes during jobs, and explain the possibilities presented by this model.

Evaluating Hadoop for Data-Intensive Scientific Operations

Emerging sensor networks, more capable instruments, and ever increasing simulation scales are generating data at a rate that exceeds our ability to effectively manage, curate, analyze, and share it. Data-intensive computing is expected to revolutionize the next-generation software stack. Hadoop, an open source implementation of the MapReduce model provides a way for large data volumes to be seamlessly processed through use of large commodity computers. The inherent parallelization, synchronization and fault-tolerance the model offers, makes it ideal for highly-parallel data-intensive applications. MapReduce and Hadoop have traditionally been used for web data processing and only recently been used for scientific applications. There is a limited understanding on the performance characteristics that scientific data intensive applications can obtain from MapReduce and Hadoop. Thus, it is important to evaluate Hadoop specifically for data-intensive scientific operations -- filter, merge and reorder-- to understand its various design considerations and performance trade-offs. In this paper, we evaluate Hadoop for these data operations in the context of High Performance Computing (HPC) environments to understand the impact of the file system, network and programming modes on performance.

Benchmarking MapReduce Implementations for Application Usage Scenarios

The MapReduce paradigm provides a scalable model for large scale data-intensive computing and associated fault-tolerance. With data production increasing daily due to ever growing application needs, scientific endeavors, and consumption, the MapReduce model and its implementations need to be further evaluated, improved, and strengthened. Several MapReduce frameworks with various degrees of conformance to the key tenets of the model are available today, each, optimized for specific features. HPC application and middleware developers must thus understand the complex dependencies between MapReduce features and their application. We present a standard benchmark suite for quantifying, comparing, and contrasting the performance of MapReduce platforms under a wide range of representative use cases. We report the performance of three different MapReduce implementations on the benchmarks, and draw conclusions about their current performance characteristics. The three platforms we chose for evaluation are the widely used Apache Hadoop implementation, Twister, which has been discussed in the literature, and LEMO-MR, our own implementation. The performance analysis we perform also throws light on the available design decisions for future implementations, and allows Grid researchers to choose the MapReduce implementation that best suits their applications needs.

LEMO-MR: Low Overhead and Elastic MapReduce Implementation Optimized for Memory and CPU-Intensive Applications

Since its inception, MapReduce has frequently been associated with Hadoop and large-scale datasets. Its deployment at Amazon in the cloud, and its applications at Yahoo! and Face book for large-scale distributed document indexing and database building, among other tasks, have thrust MapReduce to the forefront of the data processing application domain. The applicability of the paradigm however extends far beyond its use with data intensive applications and disk based systems, and can also be brought to bear in processing small but CPU intensive distributed applications. In this work, we focus both on the performance of processing large-scale hierarchical data in distributed scientific applications, as well as the processing of smaller but demanding input sizes primarily used in diskless, and memory resident I/O systems. In this paper, we present LEMO-MR (Low overhead, Elastic, configurable for in-Memory applications, and on-Demand fault tolerance), an optimized implementation of MapReduce, for both on-disk and in-memory applications, describe its architecture and identify not only the necessary components of this model, but also trade offs and factors to be considered. We show the efficacy of our implementation in terms of potential speedup that can be achieved for representative data sets used by cloud applications. Finally, we quantify the performance gains exhibited by our MapReduce implementation over Apache Hadoop in a compute intensive environment.

Parallel and distributed approach for processing large-scale XML datasets

An emerging trend is the use of XML as the data format for many distributed scientific applications, with the size of these documents ranging from tens of megabytes to hundreds of megabytes. Our earlier benchmarking results revealed that most of the widely available XML processing toolkits do not scale well for large sized XML data. A significant transformation is necessary in the design of XML processing for scientific applications so that the overall application turn-around time is not negatively affected. We present both a parallel and distributed approach to analyze how the scalability and performance requirements of large-scale XML-based data processing can be achieved. We have adapted the Hadoop implementation to determine the threshold data sizes and computation work required per node, for a distributed solution to be effective. We also present an analysis of parallelism using our Piximal toolkit for processing large-scale XML datasets that utilizes the capabilities for parallelism that are available in the emerging multi-core architectures. Multi-core processors are expected to be widely available in research clusters and scientific desktops, and it is critical to harness the opportunities for parallelism in the middleware, instead of passing on the task to application programmers. Our parallelization approach for a multi-core node is to employ a DFA-based parser that recognizes a useful subset of the XML specification, and convert the DFA into an NFA that can be applied to an arbitrary subset of the input. Speculative NFAs are scheduled on available cores in a node to effectively utilize the processing capabilities and achieve overall performance gains. We evaluate the efficacy of this approach in terms of potential speedup that can be achieved for representative XML data sets.

MARISSA: MApReduce Implementation for Streaming Science Applications

MapReduce has since its inception been steadily gaining ground in various scientific disciplines ranging from space exploration to protein folding. The model poses a challenge for a wide range of current and legacy scientific applications for addressing their “Big Data” challenges. For example: MapRe-duces best known implementation, Apache Hadoop, only offers native support for Java applications. While Hadoop streaming supports applications compiled in a variety of languages such as C, C++, Python and FORTRAN, streaming has shown to be a less efficient MapReduce alternative in terms of performance, and effectiveness. Additionally, Hadoop streaming offers lesser options than its native counterpart, and as such offers less flexibility along with a limited array of features for scientific software. The Hadoop File System (HDFS), a central pillar of Apache Hadoop is not a POSIX compliant file system. In this paper, we present an alternative framework to Hadoop streaming to address the needs of scientific applications: MARISSA (MApReduce Implementation for Streaming Science Applications). We describe MARISSAs design and explain how it expands the scientific applications that can benefit from the MapReduce model. We also compare and explain the performance gains of MARISSA over Hadoop streaming.

Benchmarking MapReduce implementations under different application scenarios

Abstract The MapReduce paradigm provides a scalable model for large scale data intensive computing and associated fault-tolerance. Data volumes generated and processed by scientific applications are growing rapidly. Several MapReduce implementations, with various degrees of conformance to the key tenets of the model, are available today. Each of these implementations is optimized for specific features. To make the right decisions, HPC application and middleware developers must thus understand the complex dependences between MapReduce features and their application. We present a set of benchmarks for quantifying, comparing, and contrasting the performance of MapReduce implementations under a wide range of representative use cases. To demonstrate the utility of the benchmarks and to provide a snapshot of the current implementation landscape, we report the performance of three different MapReduce implementations, and draw conclusions about their current performance characteristics. The three implementations we chose for evaluation are the widely used Hadoop implementation, Twister, which has been widely discussed in the literature in the context of scientific applications, and LEMO-MR which is our own implementation. We present the performance of these three implementations and draw conclusions about their performance characteristics.

Configuring a MapReduce Framework for Dynamic and Efficient Energy Adaptation

MapReduce has become a popular framework for Big Data applications. While MapReduce has received much praise for its scalability and efficiency, it has not been thoroughly evaluated for power consumption. Our goal with this paper is to explore the possibility of scheduling in a power-efficient manner without the need for expensive power monitors on every node. We begin by considering that no cluster is truly homogeneous with respect to energy consumption. From there we develop a MapReduce framework that can evaluate the current status of each node and dynamically react to estimated power usage. Inso doing, we shift power consumption work toward more energy efficient nodes which are currently consuming less power. Our work shows that given an ideal framework configuration, certain nodes may consume only 62.3% of the dynamic power they consumed when the same framework was configured as it would be in a traditional MapReduce implementation.

MARIANE: Using MApReduce in HPC environments

Abstract MapReduce is increasingly becoming a popular programming model. However, the widely used implementation, Apache Hadoop, uses the Hadoop Distributed File System (HDFS), which is currently not directly applicable to a majority of existing HPC environments such as Teragrid and NERSC that support other distributed file systems. On such resourceful High Performance Computing (HPC) infrastructures, the MapReduce model can rarely make use of full resources, as special circumstances must be created for its adoption, or simply limited resources must be isolated to the same end. This paper not only presents a MapReduce implementation directly suitable for such environments, but also exposes the design choices for better performance gains in those settings. By leveraging inherent distributed file systems’ functions, and abstracting them away from its MapReduce framework, MARIANE (MApReduce Implementation Adapted for HPC Environments) not only allows for the use of the model in an expanding number of HPC environments, but also shows better performance in such settings. This paper identifies the components and trade-offs necessary for this model, and quantifies the performance gains exhibited by our approach in HPC environments over Apache Hadoop in a data intensive setting at the National Energy Research Scientific Computing Center (NERSC).

Scalable and Distributed Processing of Scientific XML Data

A seamless and intuitive search capability for the vast amount of datasets generated by scientific experiments is critical to ensure effective use of such data by domain specific scientists. Currently, searches on enormous XML datasets is done manually via custom scripts or by using hard-to-customize queries developed by experts in complex and disparate XML query languages. Such approaches however do not provide acceptable performance for large-scale data since they are not based on a scalable distributed solution. Furthermore, it has been shown that databases are not optimized for queries on XML data generated by scientific experiments, as term kinship, range based queries, and constraints such as conjunction and negation need to be taken into account. There exists a critical need for an easy-to-use and scalable framework, specialized for scientific data, that provides natural-language-like syntax along with accurate results. As most existing search tools are designed for exact string matching, which is not adequate for scientific needs, we believe that such a framework will enhance the productivity and quality of scientific research by the data reduction capabilities it can provide. This paper presents how the MapReduce model should be used in XML metadata indexing for scientific datasets, specifically TeraGrid Information Services and the NeXus datasets generated by the Spallation Neutron Source (SNS) scientists. We present an indexing structure that scales well for large-scale MapReduce processing. We present performance results using two MapReduce implementations, Apache Hadoop and LEMO-MR, to emphasize the flexibility and adaptability of our framework in different MapReduce environments.