Real-world Scientific Applications Research Articles

Role-shifting threads: Increasing OpenMP malleability to address load imbalance at MPI and OpenMP

This paper presents the evolution of the free agent threads for OpenMP to the new role-shifting threads model and their integration with the Dynamic Load Balancing (DLB) library. We demonstrate how free agent threads can improve resource utilization in OpenMP applications with load imbalance in their nested parallel regions. We also demonstrate how DLB efficiently manages the malleability exposed by the role-shifting threads to address load imbalance issues. We use three real-world scientific applications, one of them to demonstrate that free agents alone can improve the OpenMP model without external tools, and two other MPI+OpenMP applications, one of them with a coupling case, to illustrate the potential of the free agent threads’ malleability with an external resource manager to increase the efficiency of the system. In addition, we demonstrate that the new implementation is more usable than the former one, letting the runtime system automatically make decisions that were made by the programmer previously. All software is released open-source.

Neural message-passing for objective-based uncertainty quantification and optimal experimental design

Various real-world scientific applications involve the mathematical modeling of complex uncertain systems with numerous unknown parameters. Accurate parameter estimation is often practically infeasible in such systems, as the available training data may be insufficient and the cost of acquiring additional data may be high. In such cases, based on a Bayesian paradigm, we can design robust operators retaining the best overall performance across all possible models and design optimal experiments that can effectively reduce uncertainty to enhance the performance of such operators maximally. While objective-based uncertainty quantification (objective-UQ) based on MOCU (mean objective cost of uncertainty) provides an effective means for quantifying uncertainty in complex systems, the high computational cost of estimating MOCU has been a challenge in applying it to real-world scientific/engineering problems. In this work, we propose a novel scheme to reduce the computational cost for objective-UQ via MOCU based on a data-driven approach. We adopt a neural message-passing model for surrogate modeling, incorporating a novel axiomatic constraint loss that penalizes an increase in the estimated system uncertainty. As an illustrative example, we consider the optimal experimental design (OED) problem for uncertain Kuramoto models, where the goal is to predict the experiments that can most effectively enhance robust synchronization performance through uncertainty reduction. We show that our proposed approach can accelerate MOCU-based OED by four to five orders of magnitude, without any visible performance loss compared to the state-of-the-art. The proposed approach applies to general OED tasks, beyond the Kuramoto model.

Replication-Based Dynamic Energy-Aware Resource Provisioning for Scientific Workflows

Distributed computing services in cloud environments are easily accessible to end users. These services are delivered to end users via a subscription-based model. The “infrastructure as a service” (IaaS) cloud model is one of the best cloud environment models for running data- and computing-intensive applications. Real-world scientific applications are the best examples of data and computing intensiveness. For their implementation, scientific workflow applications need high-performance computational resources and a large volume of storage. The workflow tasks are linked based on computational and data interdependence. Considering the high volume and variety of scientific workflows (SWs), the resources of the IaaS cloud model require managing energy efficiently and without failure or loss. Therefore, in order to address the issues of power consumption and task failure for real-world SWs, this research work proposes a replication-based dynamic energy-aware resource provisioning (R-DEAR) strategy for SWs in an IaaS cloud environment. The proposed strategy, R-DEAR, is a resource- and service-provisioning strategy that implements a replication-based fault-tolerant and load-balancing mechanism. The proposed R-DEAR strategy schedules the tasks of a scientific workflow with a replication-based fault-tolerant mechanism. The proposed R-DEAR strategy also manages the power consumption of IaaS cloud resources dynamically through a load-sharing process. Simulation results show that the proposed R-DEAR strategy reduces energy consumption, execution cost, and execution time by 9%, 15%, and 18%, respectively, as compared with the existing state-of-the-art strategy.

A Multi-Branch Decoder Network Approach to Adaptive Temporal Data Selection and Reconstruction for Big Scientific Simulation Data

A key challenge in scientific simulation is that the simulation outputs often require intensive I/O and storage space to store the results for effective post hoc analysis. This article focuses on a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">quality-aware adaptive temporal data selection and reconstruction</i> problem where the goal is to adaptively select simulation data samples at certain key timesteps in situ and reconstruct the discarded samples with quality assurance during post hoc analysis. This problem is motivated by the limitation of current solutions that a significant amount of simulation data samples are either discarded or aggregated during the sampling process, leading to inaccurate modeling of the simulated phenomena. Two unique challenges exist: 1) the sampling decisions have to be made in situ and adapted to the dynamics of the complex scientific simulation data; 2) the reconstruction error must be strictly bounded to meet the application requirement. To address the above challenges, we develop <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DeepSample</i> , an error-controlled convolutional neural network framework, that jointly integrates a set of coherent multi-branch deep decoders to effectively reconstruct the simulation data with rigorous quality assurance. The results on two real-world scientific simulation applications show that DeepSample significantly outperforms other state-of-the-art methods on both sampling efficiency and reconstructed simulation data quality.

MVU-Net: a multi-view U-Net architecture for weakly supervised vortex detection

Vortex detection plays a fundamental role in turbulence research and engineering problems. However, due to the lack of a mathematically rigorous vortex definition, as well as the absence of any vortex-oriented database, both traditional and machine learning detection methods achieve only limited performance. In this paper, we develop a deep learning model for vortex detection using a weak supervision approach. In order to avoid the need for a vast amount of manual labeling work, we employ an automatic clustering approach to encode vortex-like behavior as the basis for programmatically generating large-scale, highly reliable training labels. Moreover, to speed up the clustering method, a multi-view U-Net (MVU-Net) model is proposed to approximate the clustering results using the knowledge distillation technique. A multi-view learning strategy is further applied to integrate the information across multiple variables. In addition, we propose a physics-informed loss function, which enables our model to explicitly consider the characteristics of flow fields. The results on eight real-world scientific simulation applications show that the proposed MVU-Net model significantly outperforms other state-of-the-art methods on both efficiency and accuracy.

AAQAL: A Machine Learning-Based Tool for Performance Optimization of Parallel SPMV Computations Using Block CSR

The sparse matrix–vector product (SpMV), considered one of the seven dwarfs (numerical methods of significance), is essential in high-performance real-world scientific and analytical applications requiring solution of large sparse linear equation systems, where SpMV is a key computing operation. As the sparsity patterns of sparse matrices are unknown before runtime, we used machine learning-based performance optimization of the SpMV kernel by exploiting the structure of the sparse matrices using the Block Compressed Sparse Row (BCSR) storage format. As the structure of sparse matrices varies across application domains, optimizing the block size is important for reducing the overall execution time. Manual allocation of block sizes is error prone and time consuming. Thus, we propose AAQAL, a data-driven, machine learning-based tool that automates the process of data distribution and selection of near-optimal block sizes based on the structure of the matrix. We trained and tested the tool using different machine learning methods—decision tree, random forest, gradient boosting, ridge regressor, and AdaBoost—and nearly 700 real-world matrices from 43 application domains, including computer vision, robotics, and computational fluid dynamics. AAQAL achieved 93.47% of the maximum attainable performance with a substantial difference compared to in practice manual or random selection of block sizes. This is the first attempt at exploiting matrix structure using BCSR, to select optimal block sizes for the SpMV computations using machine learning techniques.

Balancing Information with Observation Costs in Deep Reinforcement Learning

The use of reinforcement learning (RL) in scientific applications, such as materials design and automated chemistry, is increasing. A major challenge, however, lies in fact that measuring the state of the system is often costly and time consuming in scientific applications, whereas policy learning with RL requires a measurement after each time step. In this work, we make the measurement costs explicit in the form of a costed reward and propose the active-measure with costs framework that enables off-the-shelf deep RL algorithms to learn a policy for both selecting actions and determining whether or not to measure the state of the system at each time step. In this way, the agents learn to balance the need for information with the cost of information. Our results show that when trained under this regime, the Dueling DQN and PPO agents can learn optimal action policies whilst making up to 50\% fewer state measurements, and recurrent neural networks can produce a greater than 50\% reduction in measurements. We postulate the these reduction can help to lower the barrier to applying RL to real-world scientific applications.

An efficient pattern-based approach for workflow supporting large-scale science: The DagOnStar experience

Workflow engines are commonly used to orchestrate large-scale scientific computations such as, but not limited to weather, climate, natural disasters, food safety, and territorial management. However, to implement, manage, and execute real-world scientific applications in the form of workflows on multiple infrastructures (servers, clusters, cloud) remains a challenge. In this paper, we present DagOnStar (Directed Acyclic Graph OnAnything), a lightweight Python library implementing a workflow paradigm based on parallel patterns that can be executed on any combination of local machines, on-premise high performance computing clusters, containers, and cloud-based virtual infrastructures. DagOnStar is designed to minimize data movement to reduce the application storage footprint. A case study based on a real-world application is explored to illustrate the use of this novel workflow engine: a containerized weather data collection application deployed on multiple infrastructures. An experimental comparison with other state-of-the-art workflow engines shows that DagOnStar can run workflows on multiple types of infrastructure with an improvement of 50.19% in run time when using a parallel pattern with eight task-level workers.

Energy Aware Parallel Scheduling Techniques for Network-on-Chip Based Systems

Minimizing execution time, energy consumption, and network load through scheduling algorithms is challenging for multi-processor-on-chip (MPSoC) based network-on-chip (NoC) systems. MPSoC based systems are prevalent in high performance computing systems. With the increase in computing capabilities of computing hardware, application requirements have increased many folds, particularly for real world scientific applications. Scheduling large scientific workflows consisting hundreds and thousands of tasks consume significant amount of time and resources. In this article, energy aware parallel scheduling techniques are presented primarily aimed at reducing the algorithm execution time while considering network load. Experimental results reveal that the proposed parallel scheduling algorithms achieve significant reduction in execution time.

Interconnection Network Energy-Aware Workflow Scheduling Algorithm on Heterogeneous Systems

Heterogeneous systems based on multicore (CPU) and manycore (GPU) processors have been regarded as an important computing infrastructure in recent years. Large-scale computationally intensive scientific workflow applications have recently been deployed on such systems. However, improving the system performance and reducing the energy consumption under user deadline constraints remain challenging problems. In this article, we first investigate the computing node network energy consumption problem of fat-tree interconnection networks for a low communication-to-computation ratio workflow application. We then propose a heuristic list-based network energy-efficient workflow scheduling (NEEWS) algorithm including top-level task computing, task subdeadline initialization, a dynamic adjustment, and an edge data optimization communication method. Extensive simulations were conducted based on randomly generated workflow applications and two real-world scientific applications. The experiment results clearly demonstrate that our proposed workflow scheduling strategy outperforms three other algorithms in terms of energy consumption. In particular, NEEWS is extremely suitable owing to its high parallelism and low communication in large-scale scientific applications.

Hybrid scheduling for scientific workflows on hybrid clouds

Scientific workflows consist of many interdependent tasks dictated by their data dependencies. As these workflows are becoming resource-intensive in both data and computing, private clouds struggle to cope with their resource requirements. Private cloud’s limitations are claimed to be addressed by public clouds, however, the complete offloading of workflow execution to public clouds causes excessive data transfer. In this paper, we address the problem of scheduling scientific workflows on hybrid clouds aiming at cost reduction while improving execution time, with the consideration of public cloud costs, by minimizing data movements between private cloud and public cloud. To this end, we develop Hybrid Scheduling for Hybrid Clouds (HSHC) as a novel data-locality aware scheduling algorithm. HSHC adopts a hybrid approach consisting of static phase and dynamic phase. The former solves the problem of workflow scheduling, using an extended genetic algorithm, with static information of workflows and resources. The latter dynamically adapts the static schedule in response to changing execution conditions, such as locality of intermediate output data and performance degradation of tasks and resources. We evaluate HSHC with both random workflows and real-world scientific applications in execution time and cost. Experimental results compared with seven state-of-the-art algorithms demonstrate HSHC significantly reduces the cost by up to 40% and improves the execution time by up to 25%.

QTMS: A quadratic time complexity topology-aware process mapping method for large-scale parallel applications on shared HPC system

Communication exacerbates the performance for parallel applications with thousands of CPU cores and quantities of data to exchange. The high communication cost is usually attributed to the mismatch between the communication patterns of parallel applications and the physical topology graphs of the computing resources (or the underlying network topologies). The topology-aware process mapping method can usually obtain a better embedding scheme with the aim to improve communication performance. Many existing heuristic-search based mapping methods have high execution time for large-scale applications. Some low-cost graph-partitioning based mapping methods depend on that the allocated resources form a regular structure, which is usually impractical in most high performance computing systems shared by multiple users and applications. This weakens their performance. Other graph-partitioning based mapping methods come at a high cost or require users to provide the network structure information. To address these issues, a quadratic time complexity topology-aware process mapping method is presented in this paper. The experimental results show that the proposed method often achieves a better application communication performance than several state-of-the-art mapping methods on a shared HPC system, while maintaining a significantly lower execution cost. Moreover, the real-world scientific application proxies gain an execution time reduction as large as 14.60% in the 512 process-scale compared to the system default process placement on the TianHe-2 HPC systems.

Predicting compressive strength of consolidated molecular solids using computer vision and deep learning

We explore the application of computer vision and machine learning (ML) techniques to predict material properties (e.g., compressive strength) based on SEM images. We show that it's possible to train ML models to predict materials performance based on SEM images alone, demonstrating this capability on the real-world problem of predicting uniaxially compressed peak stress of consolidated molecular solids samples. Our image-based ML approach reduces mean absolute percentage error (MAPE) by an average of 24% over baselines representative of the current state-of-the-practice (i.e., domain-expert's analysis and correlation). We compared two complementary approaches to this problem: (1) a traditional ML approach, random forest (RF), using state-of-the-art computer vision features and (2) an end-to-end deep learning (DL) approach, where features are learned automatically from raw images. We demonstrate the complementarity of these approaches, showing that RF performs best in the “small data” regime in which many real-world scientific applications reside (up to 24% lower RMSE than DL), whereas DL outpaces RF in the “big data” regime, where abundant training samples are available (up to 24% lower RMSE than RF). Finally, we demonstrate that models trained using machine learning techniques are capable of discovering and utilizing informative crystal attributes previously underutilized by domain experts.

Measuring the impact of burst buffers on data-intensive scientific workflows

Science applications frequently produce and consume large volumes of data, but delivering this data to and from compute resources can be challenging, as parallel file system performance is not keeping up with compute and memory performance. To mitigate this I/O bottleneck, some systems have deployed burst buffers, but their impact on performance for real-world scientific workflow applications is still not clear. In this paper, we examine the impact of burst buffers through the remote-shared, allocatable burst buffers on the Cori system at NERSC. By running two data-intensive workflows, a high-throughput genome analysis workflow, and a subset of the SCEC high-performance CyberShake workflow, a production seismic hazard analysis workflow, we find that using burst buffers offers read and write improvements of an order of magnitude, and these improvements lead to increased job performance, and thereby increased overall workflow performance, even for long-running CPU-bound jobs.

HIGA: Harmony-inspired genetic algorithm for rack-aware energy-efficient task scheduling in cloud data centers

Reducing energy consumption in cloud data centers is one of the prime issue in the cloud community. It reduces energy related costs and increases lifespan of high performance computing resources deployed in cloud data centers and it also helps in reducing carbon emissions. Along with energy efficiency, problem of task scheduling is also one of the important problem considered in cloud data centers and it belongs to NP-class problems. With the energy consumption consideration, problem of task scheduling becomes more complex to solve. Metaheuristic algorithms are proven to generate near optimal solutions for task scheduling problem but their scheduling overhead increases vastly as the number of tasks or number of resources increases. In addition to this, metaheuristic-based scheduling algorithms lives most of time in local-optimal region of solution space and it is possible that it can finalizes the solution in the local-optimal region only. Primary motivation behind this work is from genetic algorithm itself, we have experimented with genetic algorithm in many ways to solve the problem of energy efficient task scheduling and come up with new hybrid scheme to solve it. In this work, we tackles the problem of energy efficient task scheduling on modern cloud data center architecture and proposes a novel hybrid metaheuristic scheme harmony-inspired genetic algorithm (HIGA). It also addresses issues associated with metaheuristic algorithm. HIGA combines the exploration capability of genetic algorithm and exploitation capability of harmony search by which it intelligently senses local as well as global optimal region without wasting time (iterations) in local or global optimal region and provides quick convergence. Our primary objectives in this work are to reduce makespan and computing energy and secondary objectives are to reduce the energy consumed by the resources other than computing resources and reduce execution overhead associated with scheduler. Collectively these objectives guides HIGA for better energy efficiency and performance while reducing the number of required resources (i.e. active racks). It indirectly also reduces cooling energy as we can switch off the rack components (air blower, cooling inlets) once racks become idle. Simulation analysis has been performed over independent task applications as well as real-world scientific applications like CyberShake, Epigenomics and Montage. The result clearly manifests that proposed HIGA provides by up to 33% of energy savings and 47% of improvement in application performance (makespan) that too with 39% less execution overhead.

A Hybrid Algorithm for Multi-Objective Scientific Workflow Scheduling in IaaS Cloud

With the increase in deployment of scientific workflow applications on an IaaS cloud computing environment, the distribution of workflow tasks to particular cloud instances to decrease runtime and cost has emerged as an important challenge. The cloud workflow scheduling is a well-known NP-hard problem. In this paper, we propose a new approach for multi-objective workflow scheduling in IaaS clouds offering a limited amount of instances and a flexible combination of instance types, and present a hybrid algorithm combining genetic algorithm, artificial bee colony optimization and decoding heuristic for scheduling workflow tasks over the available cloud resources while trying to optimize the workflow makespan and cost simultaneously. The proposed algorithm is evaluated for real-world scientific applications by a simulation process. The simulation results show that our proposed scheduling algorithm performs better than the current state-of-the-art algorithms. We validate the results by the Wilcoxon signed-rank test.

Cost and Energy Aware Scheduling Algorithm for Scientific Workflows with Deadline Constraint in Clouds

Cloud computing is a suitable platform to execute the deadline-constrained scientific workflows which are typical big data applications and often require many hours to finish. Moreover, the problem of energy consumption has become one of the major concerns in clouds. In this paper, we present a cost and energy aware scheduling (CEAS) algorithm for cloud scheduler to minimize the execution cost of workflow and reduce the energy consumption while meeting the deadline constraint. The CEAS algorithm consists of five sub-algorithms. First, we use the VM selection algorithm which applies the concept of cost utility to map tasks to their optimal virtual machine (VM) types by the sub-makespan constraint. Then, two tasks merging methods are employed to reduce execution cost and energy consumption of workflow. Further, In order to reuse the idle VM instances which have been leased, the VM reuse policy is also proposed. Finally, the scheme of slack time reclamation is utilized to save energy of leased VM instances. According to the time complexity analysis, we conclude that the time complexity of each sub-algorithm is polynomial. The CEAS algorithm is evaluated using Cloudsim and four real-world scientific workflow applications, which demonstrates that it outperforms the related well-known approaches.

Smart PSO-based secured scheduling approaches for scientific workflows in cloud computing

Owing to its manifold advantages in adapting cloud computing for real-world scientific workflow applications, we intend to use cloud computing for executing the scientific workflows. In the present work, we aim for scheduling the workflow in the scalable resources in the cloud. In general, security is a vital challenge in cloud and so we include security constraints into our optimization model. The main objective of our work is to find an optimized schedule having minimum makespan and cost and by satisfying security demand constraint. The users can submit their security demand to the cloud provider during negotiation. The workflow is initially scheduled with list-based heuristics, which is then optimized by Particle Swarm Optimization (PSO). Thus we device a Smart Particle Swarm Optimization (SPSO)-based secured scheduling to find the optimized schedule with minimum makespan and cost. The proposed method is capable of assigning the task in the scientific workflows to the best suitable virtual machine in the cloud. Hence, the resource allocation is addressed as well by our method. Besides, a variant of PSO algorithm called Variable Neighbourhood PSO is also experimented to overcome the local optima problem. Our experimental results show that the scheduled workflows with assured security are yielding better makespan than existing methods with minimum iterations, which is well suited for cloud environment.

Reproducibility report: Team SegFAUlt @ SCC 2016

Abstract This publication is based on the reproducibility task of the 2016 Student Cluster Competition (SCC) at the Supercomputing Conference (SC) 2016, Salt Lake City, Utah, USA. The SCC is a 48 h non-stop supercomputing challenge, where teams from all over the world, consisting of six undergraduate students each, run and optimize real-world scientific workloads and applications. For the first time in 2016, one of the tasks was to reproduce the results of a scientific paper. Besides the software part of the competition the students also have to decide on a suitable hardware configuration and assemble their cluster on-site. The only limitation is the power limit of 3120 W. The reproducibility task, which is described in this publication, was based on the 2015 SC paper “A parallel connectivity algorithm for de Bruijn graphs in metagenomic applications” by Patrick Flick et al. presenting the first non-lossy parallel decomposition of metagenomic assembly. The students had to (i) report general information about the provided datasets, (ii) create similar graphs about communication and computation time to those in the original paper using their own results, (iii) compare them to conclusion of the thesis being reproduced and (iv) perform a strong scaling study on the dataset, including the generation of graphs comparable to the original ones.

An Improved Multi-Objective Optimization for Workflow Scheduling in Cloud Platform

Cloud computing is a promising research domain used to enhance the performance of information systems and applications used at our daily life. Workflow scheduling in cloud platform is a complex optimization problem because of its heterogeneous nature, complex billing models and infinite resources requirements which has lead the users to look for an optimal or a suboptimal solution that satisfy multiple competing goals such as minimum makespan and energy efficient solutions. The focus of this paper is to schedule workflow applications, optimizing makespan and energy consumption. A high-performance method for scheduling workflow on cloud computing platform will be presented in this paper, called non-dominated sorting particle swarm optimization (NSPSO). The proposed method is evaluated on synthetic representation of real world scientific workflow applications. The simulation results shows that the proposed method can provide better workflow scheduling solution for cloud when compared with other state-of-the-art methods.