Machine learning (ML) training algorithms often possess an inherent self-correcting behavior due to their iterative- convergent nature. Recent systems exploit this property to achieve adaptability and efficiency in unreliable computing environments by relaxing the consistency of execution and allowing calculation errors to be self-corrected during training. However, the behavior of such systems are only well understood for specific types of calculation errors, such as those caused by staleness, reduced precision, or asynchronicity, and for specific algorithms, such as stochastic gradient descent. In this paper, we develop a general framework to quantify the effects of calculation errors on iterative-convergent algorithms. We then use this framework to derive a worst-case upper bound on the cost of arbitrary perturbations to model parameters during training and to design new strategies for checkpoint-based fault tolerance. Our system, SCAR, can reduce the cost of partial failures by 78%{–}95% when compared with traditional checkpoint-based fault tolerance across a variety of ML models and training algorithms, providing near-optimal performance in recovering from failures.

Data analytics is undergoing a revolution in many scientific domains, and demands cost-effective parallel data analysis techniques. Traditional Java-based Big Data processing tools like Hadoop MapReduce are designed for commodity CPUs. In contrast, emerging manycore processors like the Xeon Phi have an order of magnitude greater computation power and memory bandwidth. To harness their computing capabilities, we propose the Harp-DAAL framework. We show that enhanced versions of MapReduce can be replaced by Harp, a Hadoop plug-in, that offers useful data abstractions for both high-performance iterative computation and MPI-quality communication, as well as drive Intel's native DAAL library. We select a subset of three machine learning algorithms and implement them within Harp-DAAL. Our scalability benchmarks ran on Knights Landing (KNL) clusters and achieved up to 2.5 times speedup of performance over the HPC solution in NOMAD and 15 to 40 times speedup over Java-based solutions in Spark. We further quantify the workloads on single node KNL with a performance breakdown at the micro-architecture level.

Latent Dirichlet Allocation (LDA) is a widely used machine learning technique in topic modeling and data analysis. Training large LDA models on big datasets involves dynamic and irregular computation patterns and is a major challenge to both algorithm optimization and system design. In this paper, we present a comprehensive benchmarking of our novel synchronized LDA training system HarpLDA+ based on Hadoop and Java. It demonstrates impressive performance when compared to three other MPI/C++ based state-of-the-art systems, which are LightLDA, F+NomadLDA, and WarpLDA. HarpLDA+ uses optimized collective communication with a timer control for load balance, leading to stable scalability in both shared-memory and distributed systems. We demonstrate in the experiments that HarpLDA+ is effective in reducing synchronization and communication overhead and outperforms the other three LDA training systems.

Big data processing tools have evolved rapidly in recent years. MapReduce has proven very successful but is not optimized for many important analytics, especially those involving iteration. In this regard, Iterative MapReduce frameworks improve performance of MapReduce job chains through caching. Further, Pregel, Giraph and Graph Lab abstract data as a graph and process it in iterations. But all these tools are designed with a fixed data abstraction and have limited collective communication support to synchronize application data and algorithm control states among parallel processes. In this paper, we introduce a collective communication abstraction layer which provides efficient collective communication operations on several common data abstractions such as arrays, key-values and graphs, and define a Map Collective programming model which serves the diverse collective communication demands in different parallel algorithms. We implement a library called Harp to provide the features above and plug it into Hadoop so that applications abstracted in Map Collective model can be easily developed on top of MapReduce framework and conveniently integrated with other tools in Apache Big Data Stack. With improved expressiveness in the abstraction and excellent performance on the implementation, we can simultaneously support various applications from HPC to Cloud systems together with high performance.

Large-scale iterative computations are common in many important data mining and machine learning algorithms. Most of these applications can be specified as iterations of MapReduce computations, leading to the Iterative MapReduce programming model [1] for efficient execution of data-intensive iterative computations interoperably between HPC and cloud environments. We observe that a systematic approach to collective communication is essential but notably missing in the current model. Thus we generalize the iterative MapReduce concept to Map-Collective on the premise that large collectives are a distinctive feature of data intensive and data mining applications. To show the necessity of Map-Collective model, this paper studies the implications of large-scale social image clustering problems, where 10--100 million images represented as points in a high dimensional (up to 2048) vector space are required to be divided into 1--10 million clusters.

Many scientific applications suffer from the lack of a unified approach to support the management and efficient processing of large-scale data. The Twister MapReduce Framework, which not only supports the traditional MapReduce programming model but also extends it by allowing iterations, addresses these problems. This paper describes how Twister is applied to several kinds of scientific applications such as BLAST, MDS Interpolation and GTM Interpolation in a non-iterative style and to MDS without interpolation in an iterative style. The results show the applicability of Twister to data parallel and EM algorithms with small overhead and increased efficiency.

MapReduce programming model has simplified the implementation of many data parallel applications. The simplicity of the programming model and the quality of services provided by many implementations of MapReduce attract a lot of enthusiasm among distributed computing communities. From the years of experience in applying MapReduce to various scientific applications we identified a set of extensions to the programming model and improvements to its architecture that will expand the applicability of MapReduce to more classes of applications. In this paper, we present the programming model and the architecture of Twister an enhanced MapReduce runtime that supports iterative MapReduce computations efficiently. We also show performance comparisons of Twister with other similar runtimes such as Hadoop and DryadLINQ for large scale data parallel applications.

BACKGROUND: Clouds and MapReduce have shown themselves to be a broadly useful approach to scientific computing especially for parallel data intensive applications. However they have limited applicability to some areas such as data mining because MapReduce has poor performance on problems with an iterative structure present in the linear algebra that underlies much data analysis. Such problems can be run efficiently on clusters using MPI leading to a hybrid cloud and cluster environment. This motivates the design and implementation of an open source Iterative MapReduce system Twister. RESULTS: Comparisons of Amazon, Azure, and traditional Linux and Windows environments on common applications have shown encouraging performance and usability comparisons in several important non iterative cases. These are linked to MPI applications for final stages of the data analysis. Further we have released the open source Twister Iterative MapReduce and benchmarked it against basic MapReduce (Hadoop) and MPI in information retrieval and life sciences applications. CONCLUSIONS: The hybrid cloud (MapReduce) and cluster (MPI) approach offers an attractive production environment while Twister promises a uniform programming environment for many Life Sciences applications. METHODS: We used commercial clouds Amazon and Azure and the NSF resource FutureGrid to perform detailed comparisons and evaluations of different approaches to data intensive computing. Several applications were developed in MPI, MapReduce and Twister in these different environments.