Search

Data centers and cloud environments have recently started providing graphic processing unit (GPU)-based infrastructure services. Actual general purpose GPU (GPGPU) applications have low GPU utilization, unlike GPU-friendly applications. To improve the resource utilization of GPUs, there is the need for the concurrent execution of different applications while sharing resources in a streaming multiprocessor (SM). However, it is difficult to predict the execution performance of applications because resource contention can be caused by intra-SM multitasking. Furthermore, it is crucial to find the best resource partitioning and an execution set of applications that show the best performance among many applications. To address this, the current paper proposes K-Scheduler, a multitasking placement scheduler based on the intra-SM resource-use characteristics of applications. First, the resource-use and multitasking characteristics of applications are analyzed according to their classification and their individual execution characteristics. Rules for concurrent execution are derived according to each observation, and scheduling is performed according to the corresponding rules. The results verified that the total workload execution performance of K-Scheduler improved by 18% compared to previous studies, and individual execution performance improved by 32%.

Two-stage meta-analysis has been popularly used in epidemiological studies to investigate an association between environmental exposure and health response by analyzing time-series data collected from multiple locations. The first stage estimates the location-specific association, while the second stage pools the associations across locations. The second stage often incorporates location-specific predictors (i.e., meta-predictors) to explain the between-location heterogeneity and is called meta-regression. The existing second-stage meta-regression relies on parametric assumptions and does not accommodate functional meta-predictors and spatial dependency. Motivated by these limitations, our research proposes a nonparametric Bayesian meta-regression which relaxes parametric assumptions and incorporates functional meta-predictors and spatial dependency. The proposed meta-regression is formulated by jointly modeling the association parameters and the functional meta-predictors using Dirichlet process (DP) or local DP mixtures. In doing so, the functional meta-predictors are represented parsimoniously by the coefficients of the orthonormal basis. The proposed models were applied to (1) a temperature–mortality association study and (2) suicide seasonality study, and validated through a simulation study. Supplementary materials accompanying this paper appear online.

The low capacity of main memory has become a critical issue in the performance of systems. Several memory schemes, utilizing multiple classes of memory devices, are used to mitigate the problem; hiding the small capacity by placing data in proper memory devices based on the hotness of the data. Memory tracers can provide such hotness information, but existing tracing tools incur extremely high overhead and the overhead increases as the problem size of a workload grows. In this paper, we propose Daptrace built for tracing memory access with bounded and light overhead. The two main techniques, region-based sampling and adaptive region construction, are utilized to maintain a low overhead regardless of the program size. For evaluation, we trace a wide range of 20 workloads and compared with baseline. The results show that Daptrace has a very small amount of runtime overhead and storage space overhead (1.95% and 5.38 MB on average) while maintaining the tracing quality regardless of the working set size of a workload. Also, a case study on out-of-core memory management exhibits a high potential of Daptrace for optimal data management. From the evaluation results, we can conclude that Daptrace shows great performance on identifying hot memory objects.

Graphics Processing Units (GPU) are widely used for high-speed processes in the computational science areas of biology, chemistry, meteorology, etc. and the machine learning areas of image and video analysis. Recently, data centers and cloud companies have adopted GPUs to provide them as computing resources. Because the majority of cloud providers allocate the GPU resource to users in an exclusive access method, the allocated GPU resource may not be all used. Although the method of allocating a GPU resource to multiple users for sharing can increase the resource utilization, performance degradation may occur in individual jobs because of interference between different jobs. It is difficult for a cloud provider to predict or control the performance of various applications executed on various cloud resources by considering their characteristics heuristically. Therefore, an intelligent job placement technique is required to minimize the interference between different jobs and increase resource utilization. This study defines the resource utilization history of applications and proposes a reinforcement learning-based job placement technique, which uses it as an input. For resource utilization history learning, a deep reinforcement learning model (DQN) is used. As a result of learning, the current resource’s state is not exceeded, and the resource is still provided by predicting which commonly placed jobs will have less impact on the total performance when executed simultaneously. This approach prevents the performance degradation of applications with diverse execution characteristics and increases the resource utilization by executing the applications while sharing the resources. The superiority of this study is demonstrated by using the proposed learning method and other methods to analyze workloads with various resource utilization characteristics. Through the experiments, it is proven that the proposed method facilitates a reduction of the total execution time and the effective use of resources, while the maintaining performance.

Popular deep learning frameworks like PyTorch utilize GPUs heavily for training, and suffer from out-of-memory (OOM) problems if memory is not managed properly. In this paper, we propose a modification that utilizes CUDA Unified Memory (UM) to expand GPU memory to the available host memory space so that practicality for the programmer can increase, and OOM memory errors will not result for any workload. We also pinpoint performance issues that result from our modifications to the framework, and outline future plans like reducing redundant memory copies, prefetching, and memory advising techniques to improve upon our design. Our implementation shows that PyTorch UM performance overheads are minimal when the data footprint is below GPU memory capacity.

Container based virtualization is an innovative technology that accelerates software development by providing portability and maintainability of applications. Recently, a growing number of workloads such as high performance computing (HPC) and Deep Learning(DL) are deployed in the container based environment. However, GPU resource management issues especially the GPU memory over subscription issue in container-based clusters, which brings substantial performance loss, is still challenging. This paper proposes an adaptive fair-share method to share effectively in container-based virtualization environment as well as an execution rescheduling method to manage the execution order of each container for acquiring maximum performance gain. We also proposed a checkpoint based mechanism especially for DL workload running with TensorFlow, which can efficiently solve the GPU memory over subscription problem. We demonstrate that our approach contributes to overall performance improvement as well as higher resource utilization compared to default and static fair-share methods with homogeneous and heterogeneous workloads. Compared to two other conditions, their results show that the proposed method reduces by 16.37%, 15.61% in average execution time and boosts approximately by 52.46%, 10.3% in average GPU memory utilization, respectively. We also evaluated our checkpoint based mechanism by running multiple CNN workloads with TensorFlow at the same time and the result shows our proposed mechanism can ensure each workload executing safely without out of memory (OOM) error occurs.

Storage systems in general and databases in particular usually balance between write durability and performance. It is not uncommon that write durability often relies on transaction systems that also offer a relaxed model of durability for performance. As hardware vendors provision more cores and faster storage devices, attaining fast data durability for concurrent file writes is demanding to high-performance storage systems in large-scale cluster systems. We approach the challenge by proposing a system that uses a small amount of fast persistent memory for buffering concurrent file writes while preserving data durability. The main technical issue in designing a durable file buffer is allowing concurrent file writes to store data in a shared and limited space of persistent memory without incurring lock or resource contention. This article addresses such issue and presents AutoBahn, a durable file buffer that expedites file I/O operations. To prove practicality and effectiveness, we implemented a prototype of AutoBahn in Linux-4.8.7 and ran several key-value systems—Redis, RocksDB, and WiredTiger—on AutoBahn. Evaluation results on a multicore server demonstrate that all the key-value systems achieved performance levels almost matching the non-durable counterpart. AutoBahn is a useful method that can efficiently deal with concurrent file I/O streams on multicores and fast storage devices.

Non-volatile random access memory (NVRAM) is a promising approach to persistent data storage with outstanding advantages over traditional storage devices, such as hard disk drives (HDDs) and solid state drives (SSDs). Some of its biggest advantages are its DRAM-like read latency and microsecond-level write latency, which are several hundred times faster than those in the original block device. However, one of the issues with using NVRAM as a storage device is designing an indexing system for its data stores to fully utilize NVRAM characteristics. The state-of-the-art indexing systems of non-volatile key-value stores are usually based on B+-trees or their variants, which were originally designed for block-based storage devices with better sequential performance than random performance. The semantics of B+-tree require data being sorted into leaf nodes and inner nodes and frequent splitting and merging to keep balanced. However, all the sorting, splitting, and merging operations cause extra write to NVRAM, which decreases its performance. In this article, we propose NV-Skiplist, a skiplist-based indexing system for key-value stores on NVRAM that fully uses the features of both NVRAM and DRAM. NV-Skiplist constructs its bottom layer in non-volatile memory to maintain data persistence and support range scans. It builds its upper layers in DRAM to retain rapid index searching and prevent consistently large overhead. We also propose a multiranged variant of NV-Skiplist to increase its search performance and scalability. We evaluate the performance of NV-Skiplist and wB+-tree which is a state-of-art scheme on an NVRAM emulator on a server with an Intel Xeon E5-2620 v2 processor. The results show that our design outperforms the original tree-based, non-volatile key-value stores up to 48%.

RDMA is increasingly becoming popular not only in HPC but also in data centers where high throughput and low latency are critical requirements. RDMA supports several types of transports, each of which has different characteristics, so that users can choose the right one to meet their requirements. Reliable connected (RC) transport has advantages on usability but disadvantages on scalability while unreliable datagram (UD) transport is scalable but hard to use. Dynamically connected (DC) transport has been newly introduced to address these limitations when using one of the existing transports while delivering both usability and scalability. However, despite all of these merits, DC transport is not yet generally adopted in the related fields due to hardware dependency. To eliminate the hardware dependency, in this paper, we design and implement SoftDC, a totally software-based DC transport. SoftDC uses the basic RDMA primitives to emulate the behavior of DC transport; UD transport for connection and RC transport for data transfer. We build and evaluate a SoftDC transport-based application to prove its effectiveness compared with RC transport-based one. Our experimental results show that our scheme has potential to provide the advantages of both RC and UD transports although our prototype has large connection establishment costs that can be amortized across a large data transfer.

Recently, improving the overall resource utilization through efficient scheduling of applications on graphic processing unit (GPU) clusters has been a concern. Traditional cluster-orchestration platforms providing GPUs exclusively for applications constrain high resource utilization. Co-execution of GPU applications is suggested to utilize limited resources. However, the co-execution of GPU applications without considering their diverse characteristics can lead to their unpredictable performances owing to interference resulting from contention and unbalanced usage of resources among applications. This paper proposes an interference-aware execution framework with Co-scheML for various GPU applications such as high performance computing (HPC), deep learning (DL) training, and DL inference. Various resource-usage characteristics of GPU applications are analyzed and profiled to identify various degrees of their application interference. As interference prediction is challenging owing to the complexity of GPU systems, an interference model is generated by applying defined GPU metrics to machine learning (ML) models. A Co-scheML scheduler deploys applications to minimize the interference using the predicted interference from the constructed model. Experimental results of our framework demonstrated that the resource utilization improved by 24%, the average job completion time (JCT) improved by 23%, and the makespan shortened by 22% on average, compared to baseline schedulers.

Recently, most cloud services use Docker container environment to provide their services. However, there are no researches to evaluate the performance of communication libraries for multi-GPU based distributed deep learning in a Docker container environment. In this paper, we propose an efficient communication architecture for multi-GPU based deep learning in a Docker container environment by evaluating the performances of various communication libraries. We compare the performances of the parameter server architecture and the All-reduce architecture, which are typical distributed deep learning architectures. Further, we analyze the performances of two separate multi-GPU resource allocation policies ― allocating a single GPU to each Docker container and allocating multiple GPUs to each Docker container. We also experiment with the scalability of collective communication by increasing the number of GPUs from one to four. Through experiments, we compare OpenMPI and MPICH, which are representative open source MPI libraries, and NCCL, which is NVIDIA’s collective communication library for the multi-GPU setting. In the parameter server architecture, we show that using CUDA-aware OpenMPI with multi-GPU per Docker container environment reduces communication latency by up to 75%. Also, we show that using NCCL in All-reduce architecture reduces communication latency by up to 93% compared to other libraries.

Recently, graphic processing unit (GPU) multitasking has become important in many platforms since an efficient GPU multitasking mechanism can enable more GPU-enabled tasks running on limited physical GPUs. However, current GPU multitasking technologies, such as NVIDIA Multi-Process Service (MPS) and Hyper-Q may not fully utilize GPU resources since they do not consider the efficient use of intra-GPU resources. In this paper, we present smCompactor, which is a fine-grained GPU multitasking framework to fully exploit intra-GPU resources for different workloads. smCompactor dispatches any particular thread blocks (TBs) of different GPU kernels to appropriate stream multiprocessors (SMs) based on our profiled results of workloads. With smCompactor, GPU resource utilization can be improved as we can run more workloads on a single GPU while their performance is maintained. The evaluation results show that smCompactor improves resource utilization in terms of the number of active SMs by up to 33% and it reduces the kernel execution time by up to 26% compared with NVIDIA MPS.

Incremental deep learning aims to learn a sequence of tasks while avoiding forgetting their knowledge. One naive approach using a deep architecture is to increase the capacity of the architecture as the number of tasks increases. However, this is followed by heavy memory consumption and makes the approach not practical. If we attempt to avoid such an issue with a fixed capacity, we encounter another challenging problem called catastrophic forgetting, which leads to a notable degradation of performance on previously learned tasks. To overcome these problems, we propose a clustering-guided incremental learning approach that can mitigate catastrophic forgetting while not increasing the capacity of an architecture. The proposed approach adopts a parameter-splitting strategy to assign a subset of parameters in an architecture for each task to prevent forgetting. It uses a clustering approach to discover the relationship between tasks by storing a few samples per task. When we learn a new task, we utilize the knowledge of the relevant tasks together with the current task to improve performance. This approach could maximize the efficiency of the approach realized in a single fixed architecture. Experimental results with a number of fine-grained datasets show that our method outperforms existing competitors.

Studies to understand the structure, functions, and electrophysiological properties of neurons have been conducted at the frontmost end of neuroscience. Such studies have led to the active development of high-performance research tools for exploring the neurobiology at the cellular and molecular level. Following this trend, research and application of plasmonics, which is a technology employed in high-sensitivity optical biosensors and high-resolution imaging, is essential for studying neurons, as plasmonic nanoprobes can be used to stimulate specific areas of cells. In this study, three plasmonic modalities were explored as tools to study neurons and their responses: (1) plasmonic sensing of neuronal activities and neuron-related chemicals; (2) performance-improved optical imaging of neurons using plasmonic enhancements; and (3) plasmonic neuromodulations. Through a detailed investigation of these plasmonic modalities and research subjects that can be combined with them, it was confirmed that plasmonic sensing, imaging, and stimulation techniques have the potential to be effectively employed for the study of neurons and understanding their specific molecular activities.

The advent of GPGPU (General-Purpose Graphic Processing Unit) containers enlarges opportunities of acceleration and easy-to-use in clouds. However, there is still lack of research on utilizing efficiently GPU resource and managing multiple applications at the same time. Co-execution of applications without understanding applications' execution characteristics may result in low performance caused by their interference problems. To solve the problem, this paper defines resource metrics that causes performance degradation when sharing resource. We calculate the degree of interference during concurrent execution of multi applications using a ML (Machine Learning) method with the metrics. The experiments show that the execution of interference aware groups improves 7% in execution time compared to non-interference aware group in overall. For a workload consisting of several applications, the overall performance was improved by 18% and 25%, respectively, when compared to SJF and random.

Recently, efficient execution of applications on Graphic Processing Unit(GPU) has emerged as a research topic to increase overall system throughput in cluster environment. As a current cluster orchestration platform using GPUs only supports an exclusive execution of an application on a GPU, the platform may not utilize resource of GPUs fully relying on application characteristics. Nonetheless, co-execution of GPU applications leads to interference coming from resource contention among applications. If diverse resource usage characteristics of GPU applications are not deliberated, unbalanced usage of computing resources and performance degradation could be induced in a GPU cluster. This study introduces Co-scheML for co-execution of various GPU applications such as High Performance Computing (HPC), Deep Learning (DL) Training, and DL Inference. Interference model is constructed by applying Machine Learning (ML) model with GPU metrics since predicting interference has a difficulty. Predicted interference is utilized and deployment of an application is determined by Co-scheML scheduler. Experimental results of the Co-ScheML strategy show that average job completion time is improved by 23%, and the makespan is shortened by 22% in average, as compared to baseline schedulers.

Issues related to operating Graphic Processing Unit (GPU) applications efficiently and improving overall system throughput in a GPU cluster environment exist. The platform may not utilize resource of a GPU fully depending on application characteristics because a conventional cluster orchestration platform using GPUs only supports a single execution of an application on a GPU. However, co-execution of GPU applications causes interference coming from resource contention among the applications. If various resource usage of GPU applications is not reflected, it could lead to an unbalanced usage of computing resources and consequently reduce performance in a GPU cluster. This study proposes interference-aware architecture, Co-scheML and evaluates case studies with it for workload execution of GPU applications such as High Performance Computing (HPC), Deep Learning (DL) Training, and DL Inference. Diverse resource usage is profiled to identify various degree of their interference of applications. Due to the difficulty of predicting the interference using these characteristics, interference model is generated by applying a Machine Learning (ML) model with defined GPU metrics. Proposed architecture predicts interference and deploys an application which is co-executed with a running application. Experimental results of case studies with Co-ScheML show that average job completion time is improved by 22%, and the makespan is shortened by 32% in average, as compared to baseline schedulers.

Missing data are uncollected data but meaningful for the statistical analysis due to clinical relevancy of the data for properly specified estimands in clinical trials. Meanwhile the efforts to prevent or minimize missing data are commonly applied in clinical trials, in practice, missing data still occurs. Choosing a statistical method for imputation that deals with missing data targeting specified estimands provides the more reliable estimates of treatment effects. We considered longitudinal clinical settings that have different degrees of missing data and treatment effects, and simulated different missing mechanisms using data from randomized, double-blind, placebo-controlled phase 3 confirmatory clinical trials of approved drugs. We compared four commonly used statistical methods to deal with missing data in clinical trials. We find that, when the data are missing not at random (MNAR) with higher missing rates, mixed model for repeated measurements (MMRM) method overestimates treatment difference. Pattern-mixture model estimates were seen to be more conservative in our studies than MMRM given MNAR assumptions, which are more realistic with missing data in clinical trials. We emphasize the importance of prevention of missing data and specifying the estimand based on trial objectives beforehand. The specified proper estimand and the proper statistical method might be key features to value the clinical trial results despite missing data.

With the development of cloud computing technology, many scientists want to perform their experiments in cloud environments. Because of the pay-per-use method, it is cost-optimal for scientists to only pay for the cloud services needed for their experiments. However, selection of suitable resources is difficult because they are composed of various characteristics. Therefore, a method of classification is needed to effectively take advantage of cloud resources. Static classification of a resource can derive inaccurate results, while scientists submit various experiment intentions and requirements. Thus, a dynamic resource-clustering method is needed to accurately determine application characteristics and scientists requirements. A cost-effective resource recommendation service is also needed. In this paper, a resource-clustering analysis, which considers application characteristics, and a cost-effective recommendation method in a hybrid cloud environment are proposed. The resource clustering analysis applies a self-organizing map and the k-means algorithm to cluster similar resources dynamically. In addition, the cost-effective resource recommendation method applies an efficiency metric based on application-aware resource clustering. Performance is verified by comparing the proposed clustering method with other studies resource classification methods. Results show that the proposed method can classify similar resource cluster reflecting application characteristics and recommend cost-effective resources.

Recent research and production environments are deploying more container technologies for the execution of HPC applications and for reproducing scientific workflows or computing environments. Research works, however, have not accounted for performance interference when executing corunning applications in containers though the absence of an efficient performance isolation layer cannot guarantee the absence of performance interference among multiple corunning applications which share resources. In this research, we propose an interference-aware scheduling method that mitigates the problem of performance interference based on applications’ I/O and CPU usage profiles. The proposed method estimates the amount of interference between various pairs of applications and coschedules them based on estimated interference. We evaluate the proposed method for both Bag-of-Tasks (BOT) scientific applications and scientific workflows and compare our method to the Weighted Mean Method. Our method improves the performance of the target scientific application by coscheduling applications with the least estimated interference ratios.

Recently, research has been conducted to assist in the processing and analysis of histopathological images using machine learning algorithms. In this study, we established machine learning-based algorithms to detect photothrombotic lesions in histological images of photothrombosis-induced rabbit brains. Six machine learning-based algorithms for binary classification were applied, and the accuracies were compared to classify normal tissues and photothrombotic lesions. The lesion classification model consisting of a 3-layered neural network with a rectified linear unit (ReLU) activation function, Xavier initialization, and Adam optimization using datasets with a unit size of [Formula: see text] pixels yielded the highest accuracy (0.975). In the validation using the tested histological images, it was confirmed that the model could identify regions where brain damage occurred due to photochemical ischemic stroke. Through the development of machine learning-based photothrombotic lesion classification models and performance comparisons, we confirmed that machine learning algorithms have the potential to be utilized in histopathology and various medical diagnostic techniques.

With the remarkable growth in cloud computing, computing resources can be manipulated on-demand in most scientific fields. This enables scientists to strategically select their experimental environment. Since it is hard to offer cloud resources in accordance with application characteristics, efficient resource provisioning methods are needed. This paper proposes an adaptive resource provisioning method using an application-aware machine learning technique that is based on the job history in heterogeneous infrastructures. The proposed resource provisioning method is built on two main concepts. First, it provides application-aware resource provisioning through the profiling of scientific application in a heterogeneous computing infrastructure. A resource provisioning model uses the resource usage profiles of scientific applications and job history data in heterogeneous computing infrastructures. In addition to the multilayer perceptron machine learning method, an error back-propagation approach is applied to analyze job history to re-learn the error of the output value. Second, it offers an adaptive resource scaling that is invoked by the availability of resource changes. An adaptive resource management method results in an enhancement of the overall balance between the performance and utilization of a system. For the experiments with the two CPU-intensive applications according to the method, a heterogeneous infrastructure comprising clusters and cloud environments is used. Experimental results indicate that the use of the proposed method can gratify user requests (cost and execution time) regarding its application and enhance resource usage effectiveness.

대규모 HPC 과학 응용의 워크로드가 전체 실행시간 동안 다양하게 변화하는 자원 요구사항을 갖게 되면서 특정 시점에 갑자기 요구사항이 증가하는(bursty) 형태가 되고 있다. 그러나 이러한 응용 워크로드를 고려하지 않고, 최대 자원 요구사항만을 반영한 가상 자원의 오버-프로비저닝은 과학 응용의 성능을 보장하지만 다른 응용이 사용할 수 없는 유휴 자원을 늘리는 문제로 남아있다. 본 논문에서는 OS-level 가상화 환경에서 응용의 자원 사용 패턴에 대한 프로파일링 데이터를 기반으로 메모리 자원 재구성 기법을 제안한다. 이는 유휴 상태의 메모리 자원을 신속하게 풀어주어 새로운 응용이 자원을 사용하여 수행할 수 있도록 한다. 본 연구에서는 경량화된 OS-level 가상화 시스템의 하나인 Docker에서 과학 워크플로우 응용을 이용하여 제안하는 알고리즘을 검증하였다. 실험을 통해 과학 응용을 실행하는 동안 컨테이너에 대한 메모리 할당 미세 조정이 전반적인 메모리 자원 활용을 향상시킬 수 있음을 보였다. 또한 응용의 메모리 사용 프로파일 데이터를 기반으로 하는 시뮬레이션 실험을 통해, 제안하는 동적 메모리 할당 기법을 사용하는 경우 대기 작업에 유휴상태의 메모리를 할당하여 전체 대기 작업의 수를 줄이고 시스템 작업 대기 시간이 줄어들었음을 보였다.

Graphics Processing Units (GPUs) are getting popularly utilized for multi-purpose applications in order to enhance highly performed parallelism of computation. As memory virtualization methods in GPU nodes are not efficiently provided to deal with diverse memory usage patterns for these applications, the success of their execution depends on exclusive and limited use of physical memory in GPU environments. Therefore, it is important to predict a pattern change of GPU memory usage during runtime execution of an application. Data provenance extracted from application characteristics, GPU runtime environments, input, and execution patterns from runtime monitoring, is defined for supporting application management to set runtime configuration and predict an experimental result, and utilize resource with co-located applications. In this paper, we define data provenance of an application on GPUs and manage data by profiling the execution of CUDA scientific applications. Data provenance management helps to predict execution patterns of other similar experiments and plan efficient resource configuration.

One common characteristic of modern workloads such as cloud, big data, and machine learning is memory intensiveness. In detail, such workloads tend to have a huge working set and low locality. Especially, the size of working sets is rapidly growing so that cannot be fully accommodated by a DRAM based main memory. Worse yet, the cloud computing systems, which has been pervasive since few decades ago, are continuously reducing the size of DRAM per CPU and encouraging memory overcommitment. Consequently, efficient and effective out-of-core memory management is becoming more important. Though a number of memory management mechanisms for such situations have proposed, manual analysis and optimization are still required for optimal performance of each workload due to the wide variety of data access patterns. However, existing tools for memory access analysis are not appropriate to be used here because those are not designed for extraction of the dynamic data access pattern of modern workloads. When those tools are used for the purpose, those incur unacceptably high overheads for unnecessarily accurate analysis results. To mitigate this situation, we introduce a tool that is designed for the purpose. Basically, the tool employs a memory access tracking technique based on page table entry access bit, which incurs only minimal overhead. It also provides a technique for an effective tradeoff between profiling overheads and accuracy of the output by dynamically adjusting number of tracking regions. By adopting the technique, this tool can control the level of overheads and output accuracy in bounded range that user specified regardless of the size of target workloads. The overhead can be lowered even enough to be used for online target workloads while still providing useful quality of the extracted data access pattern. The main contributions of this paper are: 1) introduce of the data access patterns profiler tool designed for modern memory-intensive workloads, and 2) empirical memory access pattern analysis of various realistic workloads.

Scientific data 6(1), 18 (2019). doi:10.1038/s41597-019-0010-0, We provide a detailed description of a serial femtosecond crystallography (SFX) dataset collected at the European X-ray free-electron laser facility (EuXFEL). The EuXFEL is the first high repetition rate XFEL delivering MHz X-ray pulse trains at 10 Hz. The short spacing (, Published by Nature Publ. Group, London

In this paper, we describe a hybrid MPI implementation of a discontinuous Galerkin scheme in Computational Fluid Dynamics which can utilize all the available processing units (CPU cores or GPU devices) on each computational node. We describe the optimization techniques used in our GPU implementation making it up to 74.88x faster than the single core CPU implementation in our machine environment. We also perform experiments on work partitioning between heterogeneous devices to measure the ideal load balance achieving the optimal performance in a single node consisting of heterogeneous processing units. The key problem is that CFD workloads need to allocate large amounts of both host and GPU device memory in order to compute accurate results. There exists an economic burden, not to mention additional communication overheads of simply scaling out by adding more nodes with high-end scientific GPU devices. In a micro-management perspective, workload size in each single node is also limited by its attached GPU memory capacity. To overcome this, we use ZFP, a floating-point compression algorithm to save at least 25% of data usage in our workloads, with less performance degradation than using NVIDIA UM.

Efficient data-aware methods in job scheduling, distributed storage management and data management platforms are necessary for successful execution of data-intensive applications. However, research about methods for data-intensive scientific applications are insufficient in large-scale distributed cloud and cluster computing environments and data-aware methods are becoming more complex. In this paper, we propose a Data-Locality Aware Workflow Scheduling (D-LAWS) technique and a locality-aware resource management method for data-intensive scientific workflows in HPC cloud environments. D-LAWS applies data-locality and data transfer time based on network bandwidth to scientific workflow task scheduling and balances resource utilization and parallelism of tasks at the node-level. Our method consolidates VMs and consider task parallelism by data flow during the planning of task executions of a data-intensive scientific workflow. We additionally consider more complex workflow models and data locality pertaining to the placement and transfer of data prior to task executions. We implement and validate the methods based on fairness in cloud environments. Experimental results show that, the proposed methods can improve performance and data-locality of data-intensive workflows in cloud environments.

As the rapid advancement and diversity in the computing systems, it is demanding to take the most robust scheduling algorithm that guarantee an optimized performance to execute manifold applications on large-scale heterogeneous computing environments. This paper present an adaptive application-aware job scheduling optimization strategy for large-scale high throughput computing in heterogeneous infrastructures. The proposed scheduling optimization method is built on two main concepts. First, it provides application-aware job distribution weights through empirical data in large-scale heterogeneous infrastructures. Here we adopt the concept of weight, which represent the ratio of tasks that will be computed on each resource. The weights can vary in terms of application type and are optimized until the system get in steady status. Secondly, it offers an adaptive control phase that is invoked by the weight adjustment and resource scaling feature. The feedback data from monitoring module is forwarded to the control phase in order to adjust weights and over-provisioning ratio, and result in enhancing overall balance between performances and utilization of system. The experimental evaluation with the four realistic workload patterns demonstrates that, when compared to the core-based scheme which distributes tasks in proportions of each resource's number of cores, the use of our optimization method can achieve 62 % better average throughput, 43 % shorter average queueing time, and 38 % better average utilization of the entire resources in diverse infrastructure environments by harnessing our adaptive module.

Journal of synchrotron radiation 26(3), 660 - 676 (2019). doi:10.1107/S1600577519003308, The European X-ray Free-Electron Laser (FEL) became the first operational high-repetition-rate hard X-ray FEL with first lasing in May 2017. Biological structure determination has already benefitted from the unique properties and capabilities of X-ray FELs, predominantly through the development and application of serial crystallography. The possibility of now performing such experiments at data rates more than an order of magnitude greater than previous X-ray FELs enables not only a higher rate of discovery but also new classes of experiments previously not feasible at lower data rates. One example is time-resolved experiments requiring a higher number of time steps for interpretation, or structure determination from samples with low hit rates in conventional X-ray FEL serial crystallography. Following first lasing at the European XFEL, initial commissioning and operation occurred at two scientific instruments, one of which is the Single Particles, Clusters and Biomolecules and Serial Femtosecond Crystallography (SPB/SFX) instrument. This instrument provides a photon energy range, focal spot sizes and diagnostic tools necessary for structure determination of biological specimens. The instrumentation explicitly addresses serial crystallography and the developing single particle imaging method as well as other forward-scattering and diffraction techniques. This paper describes the major science cases of SPB/SFX and its initial instrumentation – in particular its optical systems, available sample delivery methods, 2D detectors, supporting optical laser systems and key diagnostic components. The present capabilities of the instrument will be reviewed and a brief outlook of its future capabilities is also described., Published by Wiley-Blackwell, [S.l.]

Virtualization is an innovative technology that accelerates software development by providing portability and maintainability of applications. However, it often leads underperformance especially caused by overheads from managing virtual machines. To address the limitation of virtual machines, container technology has emerged to deploy and operate distributed applications without launching entire virtual machines. Unfortunately, resources contention issues in container-based clusters, bringing substantial performance loss are still challenging. This paper proposes an adaptive fair-share method to share effectively in container-based virtualization environment. In particular, we focus on enabling GPU sharing between multiple concurrent containers without lack of GPU memory. We demonstrate that our approach contributes to overall performance improvement as well as higher resource utilization compared to default and static fair-share methods with homogeneous and heterogeneous workloads. Compared to two other conditions, their results show that the proposed method reduces by 16.37%, 15.61% in average execution time and boosts approximately by 52.46%, 10.3% in average GPU memory utilization, respectively.

Data of scientific workflow applications are tend to be distributed over many data centers to be effectively stored, retrieved, and transferred among them. The result of an experiment with those data shows diverse execution performance depending on the placement of input and intermediate data which are generated during application execution. However, initial data placement strategy would not be the best plan for long running experiments because of the dynamic change of resource condition time to time. We propose an adaptive data placement strategy considering dynamic resource change for efficient data-intensive applications. The strategy consists of two stages that group the datasets in data centers during the build- time stage and dynamically clusters every time newly generated datasets repeatedly to the most appropriate data centers during runtime stage, which is based on task dependency, intense degree of data usage, and just-in-time resource availability. Just-in-time data placement coming with task execution is more efficient than the one with initialization stage of experiments in the aspect of resource utilization. Experiments show that data movement can be effectively reduced while the workflow is running

Scientists gain benefits from on-demand scalable resource provisioning, and various computing environments by using cloud computing resources for their applications. However, many cloud computing service providers offer their cloud resources according to their own policies. The descriptions of resource specification are diverse among vendors. Subsequently, it becomes difficult to find suitable cloud resources according to the characteristics of an application. Due to limited understanding of resource availability, scientists tend to choose resources used in previous experiments or over-performed resources without considering the characteristics of their applications. The need for standardized notations on diverse cloud resources without the constraints of complicated specification given by providers leads to active studies on intercloud to support interoperability in hybrid cloud environments. However, projects related to intercloud studies are limited as they are short of expertise in application characteristics. We define an intercloud resource classification and propose semantic resource recommendation based on statistical analysis to provide semantic cloud resource services for an application in hybrid cloud computing environments. The scheme proves benefits on resource availability and cost-efficiency with choosing semantically similar cloud resources using cluster analysis while considering application characteristics.

Recently, wellness has become an issue related to improvements in personal health and quality of life. Data that are accumulated daily, such as meals and momentum records, in addition to body measurement information such as body weight, BMI and blood pressure have been used to analyze the personal health data of an individual. Therefore, it has become possible to prevent potential disease and to analyze dietary or exercise patterns. In terms of food and nutrition, analyses are performed to evaluate the health status of an individual using dietary data. However, it is very difficult to process the large amount of dietary data. An analysis of dietary data includes four steps, and each step contains a series of iterative tasks that are executed over a long time. This paper proposes a problem solving environment that automates dietary data analysis, and the proposed framework increases the speed with which an experiment can be conducted.

Cloud computing provides on-demand resource provisioning and scalable resources dynamically for the efficient use of computing resources. Scientific applications recently need a very large number of loosely coupled tasks to be handled efficiently. In response, current computing environments often consist of heterogeneous resources such as cloud computing. To effectively use cloud resources, auto-scaling methods that consider diverse metrics such as CPU utilization and costs of resource usage have been studied widely. However it still remains a challenge to automatically and timely allocate resources such that deadline violation and application types are considered. In this paper, we propose auto-scaling methods that consider specific conditions such as application types, task dependency, user-defined deadlines and data transfer times within a hybrid computing infrastructure. Our hybrid computing infrastructure consists of local cluster and cloud resources using HTCaaS. We observe noticeable improvements in performance when our auto-scaling methods for bag-of-tasks and workflow applications is applied.

With the development of cloud computing technology, there are many scientists who want to perform their experiments in cloud environments. Because of the pay-per-use method, it is cost-optimal for scientists to only pay for the cloud services needed for their experiments. However, selection of suitable resources is difficult because they are composed of various characteristics. Therefore, a method of classification is needed to effectively utilize cloud resources. Static classification of a resource can derive inaccurate results, while scientists submit various experiment intentions and requirements. Thus, a dynamic resource-clustering method is needed to accurately determine application characteristics and scientists' requirements. In this paper, a resource-clustering analysis, which considers application characteristics in a hybrid cloud environment is proposed. The resource clustering analysis applies a self-organizing map and the k-means algorithm to cluster similar resources dynamically. Performance is verified by comparing the proposed clustering method with other studies' resource classification methods. Results show that the proposed method can classify similar resource cluster reflecting application characteristics.

Cloud computing platform is becoming one of the most desired environments for the emerging applications as well as traditional application domains. In this regards, there have been lots of efforts to develop and redesign the framework to deliver solutions that meet a variety of requirements including resilience, performance, and agility. This paper presents a design and implementation of iSDF, an integrated software-defined computing framework for scientific experiments. The iSDF aims at reinforcing traditional cloud-based scheme by supporting application-aware scheduling that avoids resource starvation and contentions for software-defined computing. To demonstrate the effectiveness of the iSDF, we evaluate it by performing several experiments based on three scenarios. The results show that our framework can achieve a 51% increase on resource utilization and average makespan time a 19% improvement compared to the conditions that do not include the policy. Thus, we proved that the framework can contribute to increase utilization for shared resources and to optimize overall makespan time, eventually affecting overall throughput of jobs.

The science cloud paradigm has been actively developed and investigated, but still requires a suitable model for science cloud system in order to support increasing scientific computation needs with high performance. This paper presents an effective provisioning model of science cloud, particularly for large-scale high throughput computing applications. In this model, we utilize job traces where a statistical method is applied to pick the most influential features to improve application performance. With these features, a system determines where VM is deployed (allocation) and which instance type is proper (provisioning). An adaptive evaluation step which is subsequent to the job execution enables our model to adapt to dynamical computing environments. We show performance achievements by comparing the proposed model with other policies through experiments and expect noticeable improvements on performance as well as reduction of cost from resource consumption through our model.

This paper presents a report on APNOMS 2013, which was held September 25---27, 2013 in the International Conference Center, Hiroshima, Japan. The theme of APNOMS 2013 was "Integrated Management of Network Virtualization".