Antonio Rosa

Dynamic distributed dimensional data model (D4M) database and computation system

A crucial element of large web companies is their ability to collect and analyze massive amounts of data. Tuple store databases are a key enabling technology employed by many of these companies (e.g., Google Big Table and Amazon Dynamo). Tuple stores are highly scalable and run on commodity clusters, but lack interfaces to support efficient development of mathematically based analytics. D4M (Dynamic Distributed Dimensional Data Model) has been developed to provide a mathematically rich interface to tuple stores (and structured query language “SQL” databases). D4M allows linear algebra to be readily applied to databases. Using D4M, it is possible to create composable analytics with significantly less effort than using traditional approaches. This work describes the D4M technology and its application and performance.

D4M 2.0 schema: A general purpose high performance schema for the Accumulo database

Non-traditional, relaxed consistency, triple store databases are the backbone of many web companies (e.g., Google Big Table, Amazon Dynamo, and Facebook Cassandra). The Apache Accumulo database is a high performance open source relaxed consistency database that is widely used for government applications. Obtaining the full benefits of Accumulo requires using novel schemas. The Dynamic Distributed Dimensional Data Model (D4M)[http://www.mit.edu/~kepner/D4M] provides a uniform mathematical framework based on associative arrays that encompasses both traditional (i.e., SQL) and non-traditional databases. For non-traditional databases D4M naturally leads to a general purpose schema that can be used to fully index and rapidly query every unique string in a dataset. The D4M 2.0 Schema has been applied with little or no customization to cyber, bioinformatics, scientific citation, free text, and social media data. The D4M 2.0 Schema is simple, requires minimal parsing, and achieves the highest published Accumulo ingest rates. The benefits of the D4M 2.0 Schema are independent of the D4M interface. Any interface to Accumulo can achieve these benefits by using the D4M 2.0 Schema.

Achieving 100,000,000 database inserts per second using Accumulo and D4M

The Apache Accumulo database is an open source relaxed consistency database that is widely used for government applications. Accumulo is designed to deliver high performance on unstructured data such as graphs of network data. This paper tests the performance of Accumulo using data from the Graph500 benchmark. The Dynamic Distributed Dimensional Data Model (D4M) software is used to implement the benchmark on a 216-node cluster running the MIT SuperCloud software stack. A peak performance of over 100,000,000 database inserts per second was achieved which is 100× larger than the highest previously published value for any other database. The performance scales linearly with the number of ingest clients, number of database servers, and data size. The performance was achieved by adapting several supercomputing techniques to this application: distributed arrays, domain decomposition, adaptive load balancing, and single-program-multiple-data programming.

Driving big data with big compute

Big Data (as embodied by Hadoop clusters) and Big Compute (as embodied by MPI clusters) provide unique capabilities for storing and processing large volumes of data. Hadoop clusters make distributed computing readily accessible to the Java community and MPI clusters provide high parallel efficiency for compute intensive workloads. Bringing the big data and big compute communities together is an active area of research. The LLGrid team has developed and deployed a number of technologies that aim to provide the best of both worlds. LLGrid MapReduce allows the map/reduce parallel programming model to be used quickly and efficiently in any language on any compute cluster. D4M (Dynamic Distributed Dimensional Data Model) provided a high level distributed arrays interface to the Apache Accumulo database. The accessibility of these technologies is assessed by measuring the effort to use these tools and is typically a few lines of code. The performance is assessed by measuring the insert rate into the Accumulo database. Using these tools a database insert rate of 4M inserts/second has been achieved on an 8 node cluster.

D4M: Bringing associative arrays to database engines

The ability to collect and analyze large amounts of data is a growing problem within the scientific community. The growing gap between data and users calls for innovative tools that address the challenges faced by big data volume, velocity and variety. Numerous tools exist that allow users to store, query and index these massive quantities of data. Each storage or database engine comes with the promise of dealing with complex data. Scientists and engineers who wish to use these systems often quickly find that there is no single technology that offers a panacea to the complexity of information. When using multiple technologies, however, there is significant trouble in designing the movement of information between storage and database engines to support an end-to-end application along with a steep learning curve associated with learning the nuances of each underlying technology. In this article, we present the Dynamic Distributed Dimensional Data Model (D4M) as a potential tool to unify database and storage engine operations. Previous articles on D4M have showcased the ability of D4M to interact with the popular NoSQL Accumulo database. Recently however, D4M now operates on a variety of backend storage or database engines while providing a federated look to the end user through the use of associative arrays. In order to showcase how new databases may be supported by D4M, we describe the process of building the D4M-SciDB connector and present performance of this connection.

LLSuperCloud: Sharing HPC systems for diverse rapid prototyping

The supercomputing and enterprise computing arenas come from very different lineages. However, the advent of commodity computing servers has brought the two arenas closer than they have ever been. Within enterprise computing, commodity computing servers have resulted in the development of a wide range of new cloud capabilities: elastic computing, virtualization, and data hosting. Similarly, the supercomputing community has developed new capabilities in heterogeneous, massively parallel hardware and software. Merging the benefits of enterprise clouds and supercomputing has been a challenging goal. Significant effort has been expended in trying to deploy supercomputing capabilities on cloud computing systems. These efforts have resulted in unreliable, low-performance solutions, which requires enormous expertise to maintain. LLSuperCloud provides a novel solution to the problem of merging enterprise cloud and supercomputing technology. More specifically LLSuperCloud reverses the traditional paradigm of attempting to deploy supercomputing capabilities on a cloud and instead deploys cloud capabilities on a supercomputer. The result is a system that can handle heterogeneous, massively parallel workloads while also providing high performance elastic computing, virtualization, and databases. The benefits of LLSuperCloud are highlighted using a mixed workload of C MPI, parallel MATLAB, Java, databases, and virtualized Web services.

Enabling on-demand database computing with MIT SuperCloud database management system

The MIT SuperCloud database management system allows for rapid creation and flexible execution of a variety of the latest scientific databases, including Apache Accumulo and SciDB. It is designed to permit these databases to run on a High Performance Computing Cluster (HPCC) platform as seamlessly as any other HPCC job. It ensures the seamless migration of the databases to the resources assigned by the HPCC scheduler and centralized storage of the database files when not running. It also permits snapshotting of databases to allow researchers to experiment and push the limits of the technology without concerns for data or productivity loss if the database becomes unstable.

LLMapReduce: Multi-level map-reduce for high performance data analysis

The map-reduce parallel programming model has become extremely popular in the big data community. Many big data workloads can benefit from the enhanced performance offered by supercomputers. LLMapReduce provides the familiar map-reduce parallel programming model to big data users running on a supercomputer. LLMapReduce dramatically simplifies map-reduce programming by providing simple parallel programming capability in one line of code. LLMapReduce supports all programming languages and many schedulers. LLMapReduce can work with any application without the need to modify the application. Furthermore, LLMapReduce can overcome scaling limits in the map-reduce parallel programming model via options that allow the user to switch to the more efficient single-program-multiple-data (SPMD) parallel programming model. These features allow users to reduce the computational overhead by more than 10x compared to standard map-reduce for certain applications. LLMapReduce is widely used by hundreds of users at MIT. Currently LLMapReduce works with several schedulers such as SLURM, Grid Engine and LSF.

Big Data strategies for Data Center Infrastructure management using a 3D gaming platform

High Performance Computing (HPC) is intrinsically linked to effective Data Center Infrastructure Management (DCIM). Cloud services and HPC have become key components in Department of Defense and corporate Information Technology competitive strategies in the global and commercial spaces. As a result, the reliance on consistent, reliable Data Center space is more critical than ever. The costs and complexity of providing quality DCIM are constantly being tested and evaluated by the United States Government and companies such as Google, Microsoft and Facebook. This paper will demonstrate a system where Big Data strategies and 3D gaming technology is leveraged to successfully monitor and analyze multiple HPC systems and a lights-out modular HP EcoPOD 240a Data Center on a singular platform. Big Data technology and a 3D gaming platform enables the relative real time monitoring of 5000 environmental sensors, more than 3500 IT data points and display visual analytics of the overall operating condition of the Data Center from a command center over 100 miles away. In addition, the Big Data model allows for in depth analysis of historical trends and conditions to optimize operations achieving even greater efficiencies and reliability.

Scalable system scheduling for HPC and big data

Abstract In the rapidly expanding field of parallel processing, job schedulers are the “operating systems” of modern big data architectures and supercomputing systems. Job schedulers allocate computing resources and control the execution of processes on those resources. Historically, job schedulers were the domain of supercomputers, and job schedulers were designed to run massive, long-running computations over days and weeks. More recently, big data workloads have created a need for a new class of computations consisting of many short computations taking seconds or minutes that process enormous quantities of data. For both supercomputers and big data systems, the efficiency of the job scheduler represents a fundamental limit on the efficiency of the system. Detailed measurement and modeling of the performance of schedulers are critical for maximizing the performance of a large-scale computing system. This paper presents a detailed feature analysis of 15 supercomputing and big data schedulers. For big data workloads, the scheduler latency is the most important performance characteristic of the scheduler. A theoretical model of the latency of these schedulers is developed and used to design experiments targeted at measuring scheduler latency. Detailed benchmarking of four of the most popular schedulers (Slurm, Son of Grid Engine, Mesos, and Hadoop YARN) is conducted. The theoretical model is compared with data and demonstrates that scheduler performance can be characterized by two key parameters: the marginal latency of the scheduler t s and a nonlinear exponent α s . For all four schedulers, the utilization of the computing system decreases to 90% for all four of the schedulers that were tested.