Execution Time Of Applications Research Articles

Improving the performance of physics applications in atom-based clusters with rCUDA

Traditionally, High-Performance Computing (HPC) has been associated with large power requirements. The reason was that chip makers of the processors typically employed in HPC deployments have always focused on getting the highest performance from their designs, regardless of the energy their processors may consume. Actually, for many years only heat dissipation was the real barrier for achieving higher performance, at the cost of higher energy consumption. However, a new trend has recently appeared consisting on the use of low-power processors for HPC purposes. The MontBlanc and Isambard projects are good examples of this trend. These proposals, however, do not consider the use of GPUs.In this paper we propose to use GPUs in this kind of low-power processor based HPC deployments by making use of the remote GPU virtualization mechanism. To that end, we leverage the rCUDA middleware in a hybrid cluster composed of low-power Atom-based nodes and regular Xeon-based nodes equipped with GPUs. Our experiments show that, by making use of rCUDA, the execution time of applications belonging to the physics domain is noticeably reduced, achieving a speed up of up to 140x with just one remote NVIDIA V100 GPU with respect to the execution of the same applications using 8 Atom-based nodes. Additionally, a rough energy consumption estimation reports improvements in energy demands of up to 37x.

CAPre: Code-Analysis based Prefetching for Persistent Object Stores

Data prefetching aims to improve access times to data storage systems by predicting data records that are likely to be accessed by subsequent requests and retrieving them into a memory cache before they are needed. In the case of Persistent Object Stores, previous approaches to prefetching have been based on predictions made through analysis of the store’s schema, which generates rigid predictions, or monitoring access patterns to the store while applications are executed, which introduces memory and/or computation overhead.In this paper, we present CAPre, a novel prefetching system for Persistent Object Stores based on static code analysis of object-oriented applications. CAPre generates the predictions at compile-time and does not introduce any overhead to the application execution. Moreover, CAPre is able to predict large amounts of objects that will be accessed in the near future, thus enabling the object store to perform parallel prefetching if the objects are distributed, in a much more aggressive way than in schema-based prediction algorithms. We integrate CAPre into a distributed Persistent Object Store and run a series of experiments that show that it can reduce the execution time of applications from 9% to over 50%, depending on the nature of the application and its persistent data model.

Unaligned Burst-Aware Memory Subsystem

In this paper, we present a memory subsystem for a burst that is unaligned to a block of synchronous dynamic random access memory (SDRAM) columns. In the case that a processor demands such an unaligned burst, a memory subsystem is required to rearrange SDRAM transfers in the order demanded by the processor. Our memory subsystem modifies a memory address given by a processor, then accesses SDRAMs with the modified memory address, and thus makes SDRAMs output transfers in an intermediate order. The transfers can be rearranged in the order demanded by the processor with the minimal performance degradation. Next, urgent transfers within the burst are delivered via priority service, and nonurgent transfers within the burst can be independently delivered via best-effort service. Experimental results show that the proposed memory subsystem achieves, on average, 16.6% shorter program execution time and 13.5% shorter memory latency than a conventional memory subsystem for wrapping bursts in the case that a quadcore processor concurrently executes four benchmarks. In addition, the proposed memory subsystem achieves, on average, 5.4% shorter application execution time and 3.4% shorter memory latency than a conventional memory subsystem for incrementing bursts in case of double data rate (DDR) 3 SDRAMs.

Workflow Scheduling in Cloud Computing

To develop scientific applications, workflow systems have become an efficient and easy vehicle. These systems use the resource provisioning, on-demand virtualized resources technology provided by Cloud computing. These virtualized resources are dynamically added and removed and users have to pay in accordance to the usage of these resources. The main issue that arises here is to decide allocation of resources to various tasks of workflow and their scheduling. As cost and execution time of applications is affected by the amount of allocated resources. In this paper the main focus is on strategies that handles workflow scheduling. There is an extensive study of such policies used in Cloud environment and also their thorough classification.

Feature Extraction Acceleration to Stabilize Execution Time for Real-Time Applications in Low-Cost Embedded Systems

Extracted features are widely used for image processing. Many research endeavors have been undertaken to extract significant features of fast moving images. Appropriate algorithm processing is necessary to extract features and provide features to the other modules in real time with low-cost embedded systems. The features from accelerated segment test (FAST) algorithm is renowned for feature extraction. FAST is composed of simple arithmetic operators. In this study, FAST is employed to implement the hardware accelerator in a field-programmable gate array for small embedded systems. Meanwhile, the threshold value in FAST affects the number of extracted features and the execution time. The precarious execution time makes it difficult for the system to schedule the timing of system functions and thus degrades the performance. An appropriate method is necessary to stabilize the execution time. A dynamic threshold controller in a FAST hardware accelerator is thus proposed to enable a stable execution time. A proportional integral controller composed of an adder, subtractor, and shifter is applied for low design implementation costs. The proposed approach occupies 2,263 slice flip-flops, 3,498 look-up tables, and 17 block RAMs in a Xilinx Virtex 5 FX field-programmable gate array. It requires 3.87[Formula: see text]ms for continuous 800×480 images from the KITTI benchmark.

MLC STT-MRAM-Aware Memory Subsystem for Smart Image Applications

Next-generation memories with high storage capacity, high performance, and low power consumption are being researched due to the ever-growing demand for artificial intelligence and high-definition applications. Among such future memories, a multi-level cell (MLC) spin-transfer magnetic torque random access memory (STT-MRAM) attracts considerable attention as an alternative to static or dynamic random access memories. An MCL STT-MRAM has the advantages of capacity and non-volatility, but the disadvantages of performance, power consumption, and endurance resulting from complicated resistance state transition and detection processes. In particular, such issues are exacerbated in the latest smart image applications employing block-based processing algorithms. In this paper, we propose a memory subsystem that mitigates the MLC STT-MRAM disadvantages in smart image applications. Our main idea is threefold: MLC-aware image buffer composing, block-aware pixel-to-memory mapping, and prediction-aware image-to-buffer allocating techniques that all make multi-step resistance state transition and detection processes less required. Experimental results show that the proposed memory subsystem achieves 24.5% shorter application execution time, and 96.4% lower memory power consumption than the conventional memory subsystems for industrial smart image applications. In addition, our memory subsystem increases the lifetime of MLC STT-MRAMs via 93.8% fewer multi-step resistance state transition processes.

Analytical composite performance models for Big Data applications

Recent years witnessed a steep rise in data generation and, consequently, the widespread adoption of software solutions able to support data-intensive applications. Many companies currently engage in data-intensive processes, however, fully embracing a data-driven paradigm is still cumbersome, and establishing a production-ready and fine-tuned deployment is time-consuming. This situation calls for innovative models and techniques to streamline the process of deployment configuration for Big Data applications. Moreover, many companies are using Cloud deployed clusters, which represent a cost-effective alternative to installation on premises. Accurate and fast prediction of the execution time of a Big Data application helps improving design time decisions, reduces Cloud over allocation charges, and assists budget management. In this paper, analytical models based on Stochastic Activity Networks (SANs) are proposed to accurately model the execution of Hadoop, Tez and Spark applications, i.e., the most referred frameworks to support Big Data analyses. The proposed SANs model these applications together with the underlying cluster in order to accurately estimate the execution time. We evaluate the accuracy of the proposed models over the TPC-DS industry benchmark across different configurations. Results obtained by numerically solving the SAN models show an average error of 4.5%, 5.8%, and 2.7% in estimating the execution time of MR, Tez, and Spark applications, respectively, against the data gathered from the experiments demonstrating higher accuracy compared with the state-of-the-art. Moreover, the time required to solve the proposed models is lower than the simulation time of the previously presented approaches in this area.

Troodon: A machine-learning based load-balancing application scheduler for CPU–GPU system

Heterogeneous computing machines consisting of a CPU and one or more GPUs are increasingly being used today because of their higher performance-cost ratio and lower energy consumption. To program such heterogeneous systems, OpenCL has become an industry standard due to the portability across various computing architectures. To exploit the computing capabilities of heterogeneous systems, application developers are porting their cluster and Cloud applications using OpenCL. With the increasing number of such applications, the use of shared accelerating computing devices (such as CPUs and GPUs) should be managed using an efficient load-balancing scheduling heuristic capable of reducing execution time, increasing throughput with high device utilization. Mostly, the OpenCL applications are suited (execute faster) on a specific computing device (CPU or GPU) and with varying data-sizes the speedup obtained by an application on the suitable device varies too. Applications’ mapping to computing devices without considering device suitability and obtainable speedup on a suitable device leads to sub-optimal execution time, lower throughput and load imbalance. Therefore, an application scheduler should consider both the device-suitability and speedup variation for scheduling decisions leading to a reduction in execution time and an increase in throughput. In this paper, we present a novel load-balancing scheduling heuristic named as Troodon that considers machine-learning based device-suitability model that classify OpenCL applications into either CPU suitable or GPU suitable. Moreover, a speedup predictor that predicts the amount of speedup that jobs will obtain when executed on a suitable device is also part of the Troodon. Troodon incorporates the E-OSched scheduling mechanism to map jobs on CPU and GPUs in a load balanced way. This results in reduced applications execution time, increased system throughput, and improved device utilization. We evaluate the proposed scheduler using a large number of data-parallel applications and compared with several other state-of-the-art scheduling heuristics. The experimental evaluation has demonstrated that the proposed scheduler outperformed the existing heuristics and reduced the application execution time up to 38% with increased system throughput and device utilization.

Multi-Cloud Management Strategies for Simulating IoT Applications

The Internet of Things (IoT) paradigm is closely coupled with cloud technologies, and the support for managing sensor data is one of the primary concerns of Cloud Computing. IoT-Cloud systems are widely used to manage sensors and different smart devices connected to the cloud, hence a large amount of data is generated by these things that need to be efficiently stored and processed. Simulation platforms have the advantage of enabling the investigation of complex systems without the need of purchasing and installing physical resources. In our previous work, we chose the DISSECT-CF simulator to model IoT-Cloud systems, and we also introduced provider policies to enable cost-aware policies resource management. The aim of this paper is to further extend the simulation capabilities of this tool by enabling multi-cloud management. In this paper we introduce four cloud selection strategies aimed to reduce application execution time and utilization costs. We detail our proposed method towards multi-cloud extension, and evaluate defined strategies through a meteorological application.

Performance and energy impact of OpenMP runtime configurations on power constrained systems

Energy consumption of an application has gained the same importance as execution time in current HPC systems, primarily due to the massive energy consumption of these systems and consequently substantial energy bills. This fact forced software developers as well as hardware vendors to put an extensive focus on energy consumption. This focus becomes even more crucial in power constrained systems, where the amount of power dedicated to applications is limited. In such systems, finding the optimal balance between execution time and energy consumption has a significant impact on utilizing these systems efficiently. A programming model is an essential layer where optimizations can be applied to find the optimal balance.OpenMP is a standard programming model for intra-node parallelism. In this paper, we give a detailed analysis of how to optimize the execution time and energy consumption of an OpenMP application using the OpenMP runtime environment at different power levels. We studied twenty-eight OpenMP parallel regions from five NAS Parallel Benchmark (NPB) applications. Based on the study, we show that for a given power level, a suitable selection of OpenMP runtime parameters can improve the execution time and energy consumption of a parallel region up to 67% and 72%, respectively. We also show that these fine-grained improvements resulted in up to 26% of execution time and 38% of energy consumption improvement for a given OpenMP application. We finally observed that optimizing the energy consumption behavior alone across all power levels regardless of execution time behavior can lead to an additional energy consumption improvement of up to 55%. However, this improvement comes at the expense of execution time.

Application of genetic algorithms and the cross‐entropy method in practical home energy management systems

Home energy management systems (HEMSs) are important platforms to allow consumers the use of flexibility in their consumption to optimise the total energy cost. The optimisation procedure embedded in these systems takes advantage of the nature of the existing loads and the generation equipment while complying with user preferences such as air temperature comfort configurations. The complexity in finding the best schedule for the appliances within an acceptable execution time for practical applications is leading not only to the development of different formulations for this optimisation problem, but also to the exploitation of non-deterministic optimisation methods as an alternative to traditional deterministic solvers. This study proposes the use of genetic algorithms (GAs) and the cross-entropy method (CEM) in low-power HEMS to solve a conventional mixed-integer linear programming formulation to optimise the total energy cost. Different scenarios for different countries are considered as well as different types of devices to assess the HEMS operation performance, namely, in terms of outputting fast and feasible schedules for the existing devices and systems. Simulation results in low-power HEMS show that GAs and the CEM can produce comparable solutions with the traditional deterministic solver requiring considerably less execution time.

Optimal Joint Scheduling and Cloud Offloading for Mobile Applications

Cloud offloading is an indispensable solution to supporting computationally demanding applications on resource constrained mobile devices. In this paper, we introduce the concept of wireless aware joint scheduling and computation offloading (JSCO) for multi-component applications, where an optimal decision is made on which components need to be offloaded as well as the scheduling order of these components. The JSCO approach allows for more degrees of freedom in the solution by moving away from a compiler pre-determined scheduling order for the components towards a more wireless aware scheduling order. For some component dependency graph structures, the proposed algorithm can shorten execution times by parallel processing appropriate components in the mobile and cloud. We define a net utility that trades-off the energy saved by the mobile, subject to constraints on the communication delay, overall application execution time, and component precedence ordering. The linear optimization problem is solved using real data measurements obtained from running multi-component applications on an HTC smartphone and the Amazon EC2, using WiFi for cloud offloading. The performance is further analyzed using various component dependency graph topologies and sizes. Results show that the energy saved increases with longer application runtime deadline, higher wireless rates, and smaller offload data sizes.

Security modeling and efficient computation offloading for service workflow in mobile edge computing

It is a big challenge for resource-limited mobile devices (MDs) to execute various complex and energy-consumed mobile applications. Fortunately, as a novel computing paradigm, edge computing (MEC) can provide abundant computing resources to execute all or parts of the tasks of MDs and thereby can greatly reduce the energy of MD and improve the QoS of applications. However, offloading workflow tasks to the MEC servers are liable to external security threats (e.g., snooping, alteration). In this paper, we propose a security and energy efficient computation offloading (SEECO) strategy for service workflows in MEC environment, the goal of which is to optimize the energy consumption under the risk probability and deadline constraints. First, we build a security overhead model to measure the execution time of security services. Then, we formulate the computation offloading problem by incorporating the security, energy consumption and execution time of workflow application. Finally, based on the genetic algorithm (GA), the corresponding coding strategies of SEECO are devised by considering tasks execution order and location and security services selection. Extensive experiments with the variety of workflow parameters demonstrate that SEECO strategy can achieve the security and energy efficiency for the mobile applications. ’

Tracing and Profiling Machine Learning Dataflow Applications on GPU

In this paper, we propose a profiling and tracing method for dataflow applications with GPU acceleration. Dataflow models can be represented by graphs and are widely used in many domains like signal processing or machine learning. Within the graph, the data flows along the edges, and the nodes correspond to the computing units that process the data. To accelerate the execution, some co-processing units, like GPUs, are often used for computing intensive nodes. The work in this paper aims at providing useful information about the execution of the dataflow graph on the available hardware, in order to understand and possibly improve the performance. The collected traces include low-level information about the CPU, from the Linux Kernel (system calls), as well as mid-level and high-level information respectively about intermediate libraries like CUDA, HIP or HSA, and the dataflow model. This is followed by post-mortem analysis and visualization steps in order to enhance the trace and show useful information to the user. To demonstrate the effectiveness of the method, it was evaluated for TensorFlow, a well-known machine learning library that uses a dataflow computational graph to represent the algorithms. We present a few examples of machine learning applications that can be optimized with the help of the information provided by our proposed method. For example, we reduce the execution time of a face recognition application by a factor of 5X. We suggest a better placement of the computation nodes on the available hardware components for a distributed application. Finally, we also enhance the memory management of an application to speed up the execution.

Assessing mobile applications performance and energy consumption through experiments and Stochastic models

Energy consumption, execution time, and availability are common terms in discussions on application development for mobile devices. Mobile applications executing in a mobile cloud computing (MCC) environment must consider several issues, such as Internet connections problems and CPU performance. Misconceptions during the design phase can have a significant impact on costs and time-to-market, or even make the application development unfeasible. Anticipating the best configuration for each type of application is a challenge that many developers are not prepared to tackle. In this work, we propose models to rapidly estimate execution time, availability, and energy consumption of mobile applications executing in an MCC environment. We defined a methodology to create and validate Deterministic and Stochastic Petri net (DSPN) models to evaluate these three critical metrics. The DSPNs results were compared with results obtained through experiments performed on a testbed environment. We analyzed an image processing application, regarding connections type (WLAN, WiFi, and 3G), servers type (MCC or cloudlet), and functionalities performance. Our numerical analyses indicate, for instance, that the use of a cloudlet significantly improves performance and energy efficiency. Besides, the baseline scenario took us one month to implement, while modeling and evaluation the three scenarios required less than one day. In this way, our DSPN models represent a powerful tool for mobile developers to plan efficient and cost-effective mobile applications. They allow rapidly assess execution time, availability, and energy consumption metrics to improve the quality of mobile applications.

On the support of inter-node P2P GPU memory copies in rCUDA

Although GPUs are being widely adopted in order to noticeably reduce the execution time of many applications, their use presents several side effects such as an increased acquisition cost of the cluster nodes or an increased overall energy consumption. To address these concerns, GPU virtualization frameworks could be used. These frameworks allow accelerated applications to transparently use GPUs located in cluster nodes other than the one executing the program. Furthermore, these frameworks aim to offer the same API as the NVIDIA CUDA Runtime API does, although different frameworks provide different degree of support. In general, and because of the complexity of implementing an efficient mechanism, none of the existing frameworks provides support for memory copies between remote GPUs located in different nodes.In this paper we introduce an efficient mechanism devised for addressing the support for this kind of memory copies among GPUs located in different cluster nodes. Several options are explored and analyzed, such as the use of the GPUDirect RDMA mechanism. We focus our discussion on the rCUDA remote GPU virtualization framework. Results show that is possible to implement this kind of memory copies in such an efficient way that performance is even improved with respect to the original performance attained by CUDA when GPUs located in the same cluster node are leveraged.

Towards providing middleware-level proactive resource reorganisation for elastic HPC applications in the cloud

Elasticity is one of the most important features of cloud computing, referring to the ability to add or remove resources according to the needs of the application or service. Particularly for High Performance Computing (HPC), elasticity can provide a better use of resources and also a reduction in the execution time of applications. Today, we observe the emergence of proactive initiatives to handle the elasticity and HPC duet, but they present at least one problem related to the need of a previous user experience, large processing time or completion of parameters. Concerning the aforesaid context, this paper presents ProElastic - a lightweight model that uses proactive elasticity to drive resource reorganisation decisions for HPC applications. Our idea is to explore both performance and adaptivity at middleware level in an effortless way at user perspective. The results showed performance gains and a competitive cost (application time × consumed resources).

GPU Static Modeling Using PTX and Deep Structured Learning

In the quest for exascale computing, energy-efficiency is a fundamental goal in high-performance computing systems, typically achieved via dynamic voltage and frequency scaling (DVFS). However, this type of mechanism relies on having accurate methods of predicting the performance and power/energy consumption of such systems. Unlike previous works in the literature, this research focuses on creating novel GPU predictive models that do not require run-time information from the applications. The proposed models, implemented using recurrent neural networks, take into account the sequence of GPU assembly instructions (PTX) and can accurately predict changes in the execution time, power and energy consumption of applications when the frequencies of different GPU domains (core and memory) are scaled. Validated with 24 applications on GPUs from different NVIDIA microarchitectures (Turing, Volta, Pascal and Maxwell), the proposed models attain a significant accuracy. Particularly, the obtained power consumption scaling model provides an average error rate of 7.9% (Tesla T4), 6.7% (Titan V), 5.9% (Titan Xp) and 5.4% (GTX Titan X), which is comparable to state-of-the-art run-time counter-based models. When using the models to select the minimum-energy frequency configuration, significant energy savings can be attained: 8.0% (Tesla T4), 6.0% (Titan V), 29.0% (Titan Xp) and 11.5% (GTX Titan X).

Using Machine Learning Ensemble Methods to Predict Execution Time of e-Science Workflows in Heterogeneous Distributed Systems

Effective planning and optimized execution of the e-Science workflows in distributed systems, such as the Grid, need predictions of execution times of the workflows. However, predicting the execution times of e-Science workflows in heterogeneous distributed systems is a challenging job due to the complex structure of workflows, variations due to input problem-sizes, and heterogeneous and dynamic nature of the shared resources. To this end, we propose two novel workflow execution time-prediction methods based on the machine learning ensemble models. In this paper, we showcase our approach for two different real Grid environments. Our approach can effectively predict the execution time of the scientific workflow applications in the Grid for various problem sizes, Grid sites, and runtime environments. We characterized the workflow performance in the Grid using the attributes that define structure of workflow as well as the execution environment. Contrary to common ensembles, our ensemble systems employed three strong learners, which balance the weaknesses of each other by their strengths to model the workflow execution times. The proposed methods have been thoroughly evaluated for three real-world e-science workflow applications. The experimental results demonstrated that our proposed multi-model ensemble models can significantly decrease the prediction error (by 50%, on average) as compared with methods based on the radial basis function neural network, local learning, and performance templates. The proposed methods can also be applied with similar effectiveness and without any major modification for other heterogeneous distributed environments, such as the Cloud.

Dependency-Aware Computation Offloading in Mobile Edge Computing: A Reinforcement Learning Approach

Mobile edge computing (MobEC) builds an Information Technology (IT) service environment to enable cloud-computing capabilities at the edge of mobile networks. To tackle the restrictions in the battery power and computation capability of mobile devices, task offloading for using MobEC is developed and used to reduce the service latency and to ensure high service efficiency. However, most of the existing schemes only focus on one-shot offloading, while taking less into consideration the task dependency. It is urgently needed a more comprehensive and adaptive way to take both the energy constraint and the inherent dependency of tasks into account, since modern communication networks have increasingly become complicated and dynamic. To this end, in this paper, we are motivated to study the problem of dependency-aware task offloading decision in MobEC, aiming at minimizing the execution time for mobile applications with constraints on energy consumption. To solve this problem, we propose a model-free approach based on reinforcement learning (RL), i.e., a Q-learning approach that adaptively learns to optimize the offloading decision and energy consumption jointly by interacting with the network environment. Simulation results show that our RL-based approach is able to achieve significant reduction on the total execution time with comparably less energy consumption.