Similarity search is one of the most fundamental computations that are regularly performed on ever-increasing protein datasets. Scalability is of paramount importance for uncovering novel phenomena that occur at very large scales. We unleash the power of over 20,000 GPUs on the Summit system to perform all-vs-all protein similarity search on one of the largest publicly available datasets with 405 million proteins, in less than 3.5 hours, cutting the time-to-solution for many use cases from weeks. The variability of protein sequence lengths, as well as the sparsity of the space of pairwise comparisons, make this a challenging problem in distributed memory. Due to the need to construct and maintain a data structure holding indices to all other sequences, this application has a huge memory footprint that makes it hard to scale the problem sizes. We overcome this memory limitation by innovative matrix-based blocking techniques, without introducing additional load imbalance.

Summary Support vector machines (SVM) is a widely used machine learning algorithm. With the increasing amount of research data nowadays, understanding how to do efficient training is more important than ever. This article discusses the performance optimizations and benchmarks related to providing high‐performance support for SVM training. In this research, we have focused on a highly scalable gradient descent‐based approach to implementing the core SVM algorithm. In providing a scalable solution, we have designed optimized high‐performance computing and dataflow‐oriented SVM implementations. A high‐performance computing approach means the algorithm is implemented with the bulk synchronous parallel (BSP) model. In addition, we analyzed the language level optimizations and math kernel optimizations on a prominent HPC modeling programming language (C++) and dataflow modeling programming language (Java). In the experiments, we compared the performance of classic HPC models, classic dataflow models, and hybrid models designed on classic HPC and dataflow programming models. Our research illustrates a scientific approach in designing the SVM algorithm at scale in classic HPC, dataflow, and hybrid systems.

Identifying similar protein sequences is a core step in many computational biology pipelines such as detection of homologous protein sequences, generation of similarity protein graphs for downstream analysis, functional annotation, and gene location. Performance and scalability of protein similarity search have proven to be a bottleneck in many bioinformatics pipelines due to increase in cheap and abundant sequencing data. This work presents a new distributed-memory software PASTIS. PASTIS relies on sparse matrix computations for efficient identification of possibly similar proteins. We use distributed sparse matrices for scalability and show that the sparse matrix infrastructure is a great fit for protein similarity search when coupled with a fully-distributed dictionary of sequences that allow remote sequence requests to be fulfilled. Our algorithm incorporates the unique bias in amino acid sequence substitution in search without altering basic sparse matrix model, and in turn, achieves ideal scaling up to millions of protein sequences.

In anagram games, players are provided with letters for forming as many words as possible over a specified time duration. Anagram games have been used in controlled experiments to study problems such as collective identity, effects of goal-setting, internal-external attributions, test anxiety, and others. The majority of work on anagram games involves individual players. Recently, work has expanded to group anagram games where players cooperate by sharing letters. In this work, we analyze experimental data from online social networked experiments of group anagram games. We develop mechanistic and data-driven models of human decision-making to predict detailed game player actions (e.g., what word to form next). With these results, we develop a composite agent-based modeling and simulation platform that incorporates the models from data analysis. We compare model predictions against experimental data, which enables us to provide explanations of human decision-making and behavior. Finally, we provide illustrative case studies using agent-based simulations to demonstrate the efficacy of models to provide insights that are beyond those from experiments alone.

To take full advantage of modern high performance architectures, many large-scale data-driven applications require loosely-coupled, fine-grained asynchronous communication. Accordingly, efficient lightweight middleware based on state-of-the-art networking technology such as RDMA is becoming a necessity. The performance critical task of handling RDMA synchronization, for existing message passing runtimes are mostly coarse granular in nature and thus may be associated with various hidden costs. While low-level RDMA libraries expose fine-grained RDMA communication that can be efficiently controlled, the critical tasks of RDMA buffer management, synchronization and flow control are left to the userspace applications requiring tedious programming effort that may lead to sub-optimal performance. In this paper we present a user-space RDMA transport layer that allows RDMA-enabled memory to be managed internally while still exposing zero-copy completion event-based RDMA transfers for message passing. The integration of an RDMA transport layer enables the opportunity for parallel applications to utilize RDMA-managed buffers for accelerating communication while co-existing with high-level MPI, GASNet or similar middleware. We show a performance speedup of up to 8X in latency/bandwidth benchmarks and 5%-90% improvement in response time or messaging rate in three reference applications with regard to their MPI implementations.

Abduction is an inference approach that uses data and observations to identify plausible (and preferably, best) explanations for phenomena. Applications of abduction (e.g., robotics, genetics, image understanding) have largely been devoid of human behavior. Here, we devise and execute an iterative abductive analysis process that is driven by the social sciences: behaviors and interactions among groups of human subjects. One goal is to understand intra-group cooperation and its effect on fostering collective identity. We build an online game platform; perform and analyze controlled laboratory experiments; form hypotheses; build, exercise, and evaluate network-based agent-based models; and evaluate the hypotheses in multiple abductive iterations, improving our understanding as the process unfolds. While the experimental results are of interest, the paper's thrust is methodological, and indeed establishes the potential of iterative abductive looping for the (computational) social sciences.

We focus on two classes of problems in graph mining: (1) finding trees and (2) anomaly detection in complex networks using scan statistics. These are fundamental problems in a broad class of applications. Most of the parallel algorithms for such problems are either based on heuristics, which do not scale very well, or use techniques like color coding, which have a high memory overhead. In this paper, we develop a novel approach for parallelizing both these classes of problems, using an algebraic representation of subgraphs as monomials-this methodology involves detecting multilinear terms in multivariate polynomials. Our algorithms show good scaling over a large regime, and they run on networks with close to half one billion edges. The resulting parallel algorithm for trees is able to scale to subgraphs of size 18, which has not been done before, and it significantly outperforms the best prior color coding based method (FASCIA) by more than two orders of magnitude. Our algorithm for network scan statistics is the first such parallelization, and it is able to handle a broad class of scan statistics functions with the same approach.

Graph scan statistics have become popular for event detection in networks. This methodology involves finding connected subgraphs that maximize a certain anomaly function, but maximizing these functions is computationally hard in general. We develop a novel approach for graph scan statistics with connectivity constraints. Our algorithm Approx-MultilinearScan relies on an algebraic technique called multilinear detection, and it improves over prior methods for large networks. We also develop a Pregel-based parallel version of this algorithm in Giraph, MultilinearScanGiraph, that allows us to solve instances with over 40 million edges, which is more than one order of magnitude larger than existing methods.

The growing use of Big Data frameworks on large machines highlights the importance of performance issues and the value of High Performance Computing (HPC) technology. This paper looks carefully at three major frameworks Spark, Flink and Message Passing Interface (MPI) both in scaling across nodes and internally over the many cores inside modern nodes. We focus on the special challenges of the Java Virtual Machine (JVM) using an Intel Haswell HPC cluster with 24 cores per node. Two parallel machine learning algorithms, K-Means clustering and Multidimensional Scaling (MDS) are used in our performance studies. We identify three major issues - thread models, affinity patterns, and communication mechanisms - as factors affecting performance by large factors and show how to optimize them so that Java can match the performance of traditional HPC languages like C. Further we suggest approaches that preserve the user interface and elegant dataflow approach of Flink and Spark but modify the runtime so that these Big Data frameworks can achieve excellent performance and realize the goals of HPC-Big Data convergence.

Large volumes of high dimensional time series data are increasingly becoming commonplace, and the ability to project such data into three dimensional space to visually inspect them is an important capability for scientific exploration. Algorithms such as Multidimensional Scaling (MDS) and Principal Component Analysis (PCA) can be used to reduce high dimensional data into a lower dimensional space. The time sensitive nature of such data requires continuous processing in time windows and visualizations to be shown as moving plots. In this paper we present: 1. an MDS-based approach to project high dimensional time series data to 3D with automatic transformation to align successive data segments; 2. an open source commodity visualization of three-dimensional time series in web browser based on Three.js; and 3. An example based on stock market data. The paper discusses various options available when producing the visualizations and how one optimizes the heuristic methods based on experimental results.

Phylogenetic analysis is commonly used to analyze genetic sequence data from fungal communities, while ordination and clustering techniques commonly are used to analyze sequence data from bacterial communities. However, few studies have attempted to link these two independent approaches. In this paper, we propose a method, which we call spherical phylogram (SP), to display the phylogenetic tree within the clustering and visualization result from a pipeline called DACIDR. In comparison with traditional tree display methods, the correlations between the tree and the clustering can be observed directly. In addition, we propose an algorithm called interpolative joining (IJ) to construct and visualize the SP in 3D space. In the experiments, we used the sum of branch lengths to quantify the general fit between the clustering and the phylogenetic tree in SP and Mantel tests to determine how well the same grouping of sequences was preserved between the clustering and the SP. Our results show that DACIDR has a classification accuracy that is similar to a phylogenetic tree generated using a multiple sequence alignment, while having much lower computational cost.

SUMMARY Modern biology is experiencing a rapid increase in data volumes that challenges our analytical skills and existing cyberinfrastructure. Exponential expansion of the protein sequence universe (PSU), the protein sequence space, together with the costs and complexities of manual curation creates a major bottleneck in life sciences research. Existing resources lack scalable visualization tools that are instrumental for functional annotation. Here, we describe a new visualization tool using multidimensional scaling to create a 3D embedding of the protein space. The advantages of the proposed PSU method include the ability to scale to large numbers of sequences, integrate different similarity measures with other functional and experimental data, and facilitate protein annotation. We applied the method to visualize the prokaryotic PSU using sequence alignment scores. As an annotation example, we used the interpolation approach to map the set of annotated archaeal proteins into the prokaryotic PSU. Transdisciplinary approaches akin to the one described in this paper are urgently needed to quickly and efficiently translate the influx of new data into tangible innovations and groundbreaking discoveries. Copyright © 2013 John Wiley & Sons, Ltd.

BACKGROUND: Modern pyrosequencing techniques make it possible to study complex bacterial populations, such as 16S rRNA, directly from environmental or clinical samples without the need for laboratory purification. Alignment of sequences across the resultant large data sets (100,000+ sequences) is of particular interest for the purpose of identifying potential gene clusters and families, but such analysis represents a daunting computational task. The aim of this work is the development of an efficient pipeline for the clustering of large sequence read sets. METHODS: Pairwise alignment techniques are used here to calculate genetic distances between sequence pairs. These methods are pleasingly parallel and have been shown to more accurately reflect accurate genetic distances in highly variable regions of rRNA genes than do traditional multiple sequence alignment (MSA) approaches. By utilizing Needleman-Wunsch (NW) pairwise alignment in conjunction with novel implementations of interpolative multidimensional scaling (MDS), we have developed an effective method for visualizing massive biosequence data sets and quickly identifying potential gene clusters. RESULTS: This study demonstrates the use of interpolative MDS to obtain clustering results that are qualitatively similar to those obtained through full MDS, but with substantial cost savings. In particular, the wall clock time required to cluster a set of 100,000 sequences has been reduced from seven hours to less than one hour through the use of interpolative MDS. CONCLUSIONS: Although work remains to be done in selecting the optimal training set size for interpolative MDS, substantial computational cost savings will allow us to cluster much larger sequence sets in the future.

We present performance results on a Windows cluster with up to 768 cores using MPI and two variants of threading - CCR and TPL. CCR (Concurrency and Coordination Runtime) presents a message based interface while TPL (Task Parallel Library) allows for loops to be automatically parallelized. MPI is used between the cluster nodes (up to 32) and either threading or MPI for parallelism on the 24 cores of each node. We use a simple matrix multiplication kernel as well as a significant bioinformatics gene clustering application. We find that the two threading models offer similar performance with MPI outperforming both at low levels of parallelism but threading much better when the grain size (problem size per process) is small. We find better performance on Intel compared to AMD on comparable 24 core systems. We develop simple models for the performance of the clustering code.

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.