Job Completion Research Articles

Towards Accelerating Data Intensive Application's Shuffle Process Using SmartNICs

The wide adoption of the emerging SmartNIC technology creates new opportunities to offload application-level computation into the networking layer, which frees the burden of host CPUs, leading to performance improvement. Shuffle, the all-to-all data exchange process, is a critical building block for network communication in distributed data-intensive applications and can potentially benefit from SmartNICs. In this paper, we develop SmartShuffle, which accelerates the data-intensive application's shuffle process by offloading various computation tasks into the SmartNIC devices. SmartShuffle supports offloading both low-level network functions, including data partitioning and network transport, and high-level computation tasks, including filtering, aggregation, and sorting. SmartShuffle adopts a coordinated offload architecture to make sender-side and receiver-side SmartNICs jointly contribute to the benefits of shuffle computation offload. SmartShuffle carefully manages the tight and time-varying computation and memory constraints on the device. We propose a liquid offloading approach, which dynamically migrates operators between the host CPU and the SmartNIC at runtime such that resources in both devices are fully utilized. We prototype SmartShuffle on the Stingray SoC SmartNICs and plug it into Spark. Our evaluation shows that SmartShuffle improves host CPU efficiency and I/O efficiency with lower job completion time. SmartShuffle outperforms Spark, and Spark RDMA by up to 40% on TPC-H.

12 Exposure to Metals and Particles in Welding and Episodes of Asthma/Wheeze and Rhinitis: a Canadian Cohort Study.

Abstract Introduction Welders are exposed through welding fume to particles and metals, including nickel and chromium, that may act as respiratory sensitizers. In a Canadian study of male and female welders we investigated the relation of modelled welding exposures to episodes of asthma/wheeze and rhinitis. Methods Welders who had been registered in an apprenticeship were recruited and followed up at 6-month intervals for up to five years. Women were recruited from across Canada, men just from Alberta. At each contact participants were asked about their health and specifically about symptoms of asthma and rhinitis and age these had first occurred. A complete job history was taken and at each contact after recruitment the participant completed a detailed welding task questionnaire. A job exposure matrix was developed, reflecting type of welding, base metal and consumables. This was validated against urinary metal concentration. Results 1001 welders, including 447 women, were recruited. In bivariate analyses, hours of welding/week, hours grinding/day and estimated exposures to total particles, manganese, nickel and aluminum were related to episodes of asthma/wheeze. Hours welding, grinding and nickel exposure were related to rhinitis. In multivariable models, including all exposures and confounders, the number of hours welding/week and hours grinding/day were related to asthma/wheeze episodes and hours grinding and nickel exposure to rhinitis. Conclusions Duration of welding tasks was related to episodes of asthma/wheeze and rhinitis. There was less evidence that estimated metal concentrations were associated with respiratory ill-health, having adjusted for hours welding.

An Algebraic Approach to the Solutions of the Open Shop Scheduling Problem

The open shop scheduling problem (OSSP) is one of the standard scheduling problems. It consists of scheduling jobs associated with a finite set of tasks developed by different machines. In this case, each machine processes at most one operation at a time, and the job processing order on the machines does not matter. The goal is to determine the completion times of the operations processed on the machines to minimize the largest job completion time, called Cmax. This paper proves that each OSSP has associated a path algebra called Brauer configuration algebra whose representation theory (particularly its dimension and the dimension of its center) can be given using the corresponding Cmax value. It has also been proved that the dimension of the centers of Brauer configuration algebras associated with OSSPs with minimal Cmax are congruent modulo the number of machines.

Cloud Configuration Optimization for Recurring Batch-Processing Applications

Recognizing the diversity of Big Data analytic jobs, cloud providers offer a wide range of VM instance types or even clusters to cater for different use cases. The choice of cloud configurations can have a significant impact on the response time and running cost of batch-processing applications, which may need to be re-run regularly with cloud-scale resources. However, identifying the best cloud configuration with a low search cost is quite challenging due to i) the large and high-dimensional configuration space, ii) the time-varying cloud service cost (e.g., AWS Spot instances), and iii) job response time variation even given the same configuration. To tackle these challenges, we design and implement <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Accordia</i> , a system that enables Adaptive Cloud Configuration Optimization for Recurring Data-Intensive Applications. By leveraging recent algorithmic advances in Gaussian Process UCB techniques, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Accordia</i> can unearth the cost-optimal configuration with a deadline constraint (i.e., maximum tolerated running time) under the time-varying cloud service cost. More importantly, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Accordia</i> manages to achieve a theoretical performance guarantee, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">sub-linearly increasing dynamic regret</i> of the job completion cost. Using extensive trace-driven simulations and empirical measurements of our Kubernetes-based implementation, we demonstrate that <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Accordia</i> can identify a near-cost-optimal configuration (i.e., within 10% of the optimum) after fewer than 20 runs from over 7000 candidate choices, which translates to a 2X-speedup and up to 17.9% cost-savings, when comparing to the state-of-the-art approach, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">CherryPick</i> .

Online scheduling of coflows by attention-empowered scalable deep reinforcement learning

With the abstraction of parallel data transmission flows being a coflow, data transmissions in large-scale computing jobs can be modeled by a coflow directed acyclic graph (coflow DAG) in which nodes are coflows and edges represent dependencies between coflows. Efficient scheduling of coflows on network links is crucial for reducing the overall communication and job completion time. The known best coflow scheduling method deploying deep reinforcement learning (DRL), DeepWeave (Sun et al., 2020), suffers from poor scalability due to the requirement of O(dn)-size policy network for processing n coflows of d dimensions which is difficult to train. This paper extends the directed acyclic graph neural network (DAGNN) to Pipelined-DAGNN that embeds the features of different stages of input coflow DAGs in pipeline to effectively speed up the feature extraction process. To effectively process the feature vectors of coflow DAGs of arbitrary size and shape without compromising scheduling accuracy (quality), we propose a novel self-attention empowered DRL coflow scheduling model to generate coflow scheduling policies, which enables the scale of policy network depends only on features (dimensions) rather than coflows, without the need of packing all individual embedding vectors from Pipelined-DAGNN into a long flat vector. Our model reduces the size of the policy network in DRL from previously O(dn) to O(d), achieving a high scalability independent of the number of coflows. Simulation results on Facebook trace show that our model reduces the average weighted job completion time by up to 33.88%, apart from being more scalable and robust, compared with the state-of-the-art methods.

DFS: Joint data formatting and sparsification for efficient communication in Distributed Machine Learning

Efficient communication is crucial to Distributed Machine Learning (DML). In this work, we propose an approach jointing Data Formatting and Sparsification (DFS) to optimize the communication in DML systems based on the parameter server framework. By doing so, we can reduce the time to transmit (aggregated) gradients between the parameter server and workers, and consequently the time to complete training jobs. More specifically, in DFS, every worker first tries to derive as many blocks with all-zero gradients as possible via sparsification, and transmits gradients block by block in a streaming fashion. By skipping blocks with all-zero gradients, we can reduce the communication cost for gradient transmission. Different from previous works on optimizing the communication in DML systems, DFS has three distinct features: (i). it dynamically determines the gradient block size; (ii). it takes into consideration both the data transfer from workers to the parameter server and that from the parameter server to workers; and (iii). it jointly optimizes the data formatting and sparsification. In other words, it performs sparsification in the way that helps form more (or larger) all-zero blocks and save more communication cost. By implementing DFS on a real testbed, we find that it can reduce the time to train a ResNet-18 model by 74.12%. Through extensive simulations, we demonstrate that DFS outperforms the state-of-the-art technique, i.e., OmniReduce (Fei et al., 2021), by up to 87.17% in terms of reducing communication cost in DML systems.

Multi-Stage Geo-Distributed Data Aggregation With Coordinated Computation and Communication in Edge Compute First Networking

With the explosion of geo-distributed data, the huge treasures hidden in them are waiting to be explored to obtain valuable insights. This results in the need for an effective geo- distributed data analysis method. The traditional approach to geo-distributed data analytics is to gather all the required data into a single edge datacenter (edge DC) through one transmission and aggregation (centralized data aggregation). However, as the volume of data grows exponentially, the centralized data aggregation scheme becomes inefficient or infeasible due to the limitations of the computing and network resources. In this paper, we propose the geo-distributed data aggregation scheme in edge compute first networking (CFN) with joint consideration of computation and communication resources. The proposed scheme optimizes two objectives: the first is to minimize the job completion time (JCT) by selecting cluster centers, dividing clusters and provisioning lightpaths; the second objective is to reduce bandwidth consumption by reallocating routing and frequency slots based on JCT. To achieve the two objectives, we first formulate the optimization problem of multi-stage geo- distributed data aggregation as a linear programming (LP) model. To tackle the computational complexity issue of the LP model, a multi-stage geo-distributed data aggregation algorithm jointly with computation and communication resources (MGDD-CC) is proposed. Simulation results show that the proposed scheme can reduce JCT, alleviate the competition for bandwidth resources and is more suitable for scenarios with better data aggregation effects and larger quantities of geo-distributed data.

The Emergence of Public Sector Innovation Associated with Civil Servants' Perception in 3T Regions: Results of a Multiple Regression Analysis

Innovations from civil servants determine government bureaucracy. However, their effectiveness toward the emergence of public sector innovations has not been widely explored, particularly from civil servants' perception of local government institutions in a 3T region. This study examines the civil servants' perception of the emergence of public sector innovations and identifies the factors that influence them. Thirty-nine civil servants of the innovation laboratory WhatsApp group from North Lombok Regency in the 3T regions participated in this study. The data were analysed using descriptive statistics and multiple regression. The result of the study discovered that most civil servants have a positive insight into the emergence of public sector innovations. They perceived that leader are essential in encouraging innovations in the public sector. Their organisational climate is open to new things. However, attention regarding employee rewards after completing work challenges is still lacking. Meanwhile, their organisation has sufficient capacity to encourage job completion. Multiple regression analysis revealed that independent variables (transformational leadership, organisational climate, and organisational capacity) have a 54.9% contribution and simultaneously affect the emergence of public sector innovations. The partial regression coefficient test revealed that organizational climate contributed the most, with a 33.4% contribution to the emergence of public sector innovations. The results of this study provide implications for both academic and government settings, especially in the organisational climate, which turned out to have been significantly affected by the emergence of public sector innovations

Online Scheduling Algorithm for Heterogeneous Distributed Machine Learning Jobs

Distributed machine learning (ML) has played a key role in todays proliferation of AI services. A typical model of distributed ML is to partition training datasets over multiple worker nodes to update model parameters in parallel, adopting a parameter server or AllReduce architecture. A fundamental problem in a distributed ML cluster is how to explore the demand elasticity of ML jobs and schedule them with different resource configurations, such that the average job completion time is minimized. To address it, we propose an online scheduling algorithm to decide the execution time window, the number and the type of concurrent workers and parameter servers for each job upon its arrival, with a goal of minimizing the weighted average completion time. Our online algorithm consists of (i) an online scheduling framework that groups unprocessed ML training jobs into a batch iteratively, and (ii) a batch scheduling algorithm that configures each ML job to maximize the total weight of scheduled jobs in the current iteration. Our online algorithm guarantees a good parameterized competitive ratio with polynomial time complexity. Extensive evaluations using real-world data demonstrate that it outperforms state-of-the-art schedulers in todays AI cloud systems.

Elastic Resource Management for Deep Learning Applications in a Container Cluster

The increasing demand for learning from massive datasets is restructuring our economy. Effective learning, however, involves nontrivial computing resources. Most businesses utilize commercial infrastructure providers (e.g., AWS) to host their computing clusters in the cloud, where various jobs compete for available resources. While cloud resource management is a fruitful research field that has made many advances in production, such as Kubernetes and YARN, few efforts have been invested to further optimize the system performance, especially for Deep Learning (DL) training jobs in a container cluster. This work introduces FlowCon, a system that is able to monitor the individual evaluation functions of DL jobs at runtime, and thus to make placement decisions and resource allocations elastically. We present a detailed design and implementation of FlowCon and conduct intensive experiments over various DL models. The results demonstrate that FlowCon significantly improves DL job completion time and resource utilization efficiency when compared to default systems. According to the results, FlowCon can improve the completion time by up to 68.8&#x0025; and meanwhile, reduce the makespan by 18.0&#x0025;, in the presence of various DL job workloads.

Deep Learning-Based Job Placement in Distributed Machine Learning Clusters With Heterogeneous Workloads

Nowadays, most leading IT companies host a variety of distributed machine learning (ML) workloads in ML clusters to support AI-driven services, such as speech recognition, machine translation, and image processing. While multiple jobs are executed concurrently in a shared cluster to improve resource utilization, interference among co-located ML jobs can lead to significant performance downgrade. Existing cluster schedulers, such as YARN and Mesos, are interference-agnostic in their job placement, leading to suboptimal resource efficiency and usage. Some literature has studied interference-aware job placement policy, but relies on detailed workload profiling and interference modeling, which is not a general solution. In this work, we present Harmony, a deep learning-driven ML cluster scheduler that places heterogeneous training jobs (either with parameter server architecture or all-reduce architecture) in a manner that minimizes interference and maximizes performance (i.e., training completion time minimization). The design of Harmony is based on a carefully designed deep reinforcement learning (DRL) framework enhanced with reward modeling. The DRL integrates a dynamic sequence-to-sequence model with the state-of-the-art techniques to stabilize training and improve convergence, including actor-critic algorithm, job-aware action space exploration, multi-head attention, and experience replay. In view of a common lack of reward samples corresponding to different placement decisions, we build an auxiliary sequence-to-sequence reward prediction model, which is trained with historical samples and used for producing reward for unseen placement. Experiments using real ML workloads in a Kubernetes cluster of 6 GPU servers show that Harmony outperforms representative schedulers by 16%–42% in terms of average job completion time.

Time-Aware Data Partition Optimization and Heterogeneous Task Scheduling Strategies in Spark Clusters

Abstract The Spark computing framework provides an efficient solution to address the major requirements of big data processing, but data partitioning and job scheduling in the Spark framework are the two major bottlenecks that limit Spark’s performance. In the Spark Shuffle phase, the data skewing problem caused by unbalanced data partitioning leads to the problem of increased job completion time. In response to the above problems, a balanced partitioning strategy for intermediate data is proposed in this article, which considers the characteristics of intermediate data, establishes a data skewing model and proposes a dynamic partitioning algorithm. In Spark heterogeneous clusters, because of the differences in node performance and task requirements, the default task scheduling algorithm cannot complete scheduling efficiently, which leads to low system task processing efficiency. In order to deal with the above problems, an efficient job scheduling strategy is proposed in this article, which integrates node performance and task requirements, and proposes a task scheduling algorithm using greedy strategy. The experimental results prove that the dynamic partitioning algorithm for intermediate data proposed in this article effectively alleviates the problem that data skew leads to the decrease of system task processing efficiency and shortens the overall task completion time. The efficient job scheduling strategy proposed in this article can efficiently complete the job scheduling tasks under heterogeneous clusters, allocate jobs to nodes in a balanced manner, decrease the overall job completion time and increase the system resource utilization.

On the use of intelligent metasurfaces in data centers

On-demand wireless links in hybrid data centers can augment the existing network capacity to improve the performance. In the absence of line of sight between the rack transceivers, reflectors have been considered in literature to support high bandwidth, on-demand wireless links within the data centers. Metasurfaces, as a candidate smart reflector technology can be harnessed to bring the intelligence burden to the reflector side which, coupled with the expected low power expenditure, can render the technology a useful alternative. Given the fact that this has not been considered in the literature before, in this paper we highlight the design challenges which emanate from such a consideration and present tools and methodologies which can be utilized to address these challenges, focusing on two problems: wireless link selection to be served by the metasurface, and workload characterization within the metasurface. The former is formulated as an optimization problem with demonstrated effectiveness in minimizing the job completion time. Moreover, the workload analysis reveals a number of interesting attributes, as for example, uneven spatial distribution of traffic on the metasurface's controller network and how this is affected by the metasurface size and location, while the use of beam splitting on the metasurface to support parallel processing and storage functionalities does not significantly affect the incurred workload. Unlike previous work on metasurfaces, which consider mobility-driven reconfigurations, in this paper for the first time we consider data-driven reconfigurations.

Increased Task Execution with a Bandwidth-Aware Hadoop Scheduling Approach

This research presents a novel bandwidth-aware Hadoop scheduling method that addresses the challenge of task scheduling in Hadoop clusters while considering the real-time network conditions. The proposed method involves the establishment of a job time completion model and a mathematical model for a Hadoop scheduling system. Furthermore, it transforms the Hadoop task scheduling problem into an optimization problem to find the task scheduling method that minimizes job completion time. By leveraging Software-Defined Networking (SDN) capabilities, a time slot-based network bandwidth allocation mechanism is introduced to allocate bandwidth fairly across network links. The proposed method also takes into account task locality and network bandwidth availability when allocating computational nodes for individual tasks. Through this approach, the limitations of existing methods, which fail to simultaneously consider global task scheduling and actual network bandwidth availability, are overcome. Experimental evaluations demonstrate the effectiveness of the proposed method in enhancing the performance of Hadoop task scheduling.

Machine Learning Feature Based Job Scheduling for Distributed Machine Learning Clusters

With the rapid proliferation of Machine Learning (ML) and Deep learning (DL) applications running on modern platforms, it is crucial to satisfy application performance requirements such as meeting deadline and ensuring accuracy. To this end, researchers have proposed several job schedulers for ML clusters. However, none of the previously proposed schedulers consider ML model parallelism, though it has been proposed as an approach to increase the efficiency of running large-scale ML and DL jobs. Thus, in this paper, we propose an ML job Feature based job Scheduling system (MLFS) for ML clusters running both data parallelism and model parallelism ML jobs. MLFS first uses a heuristic scheduling method that considers an ML job&#x2019;s spatial and temporal features to determine task priority for job queue ordering in order to improve job completion time (JCT) and accuracy performance. It uses the data from the heuristic scheduling method for training a deep reinforcement learning (RL) model. After the RL model is well trained, it then switches to the RL method to automatically make decisions on job scheduling. In addition, MLFS has a system load control method that selects tasks from overloaded servers to move to underloaded servers based on task priority, and also intelligently removes the tasks that generate little or no improvement on the desired accuracy performance when the system is overloaded to improve JCT and accuracy by job deadline. Furthermore, we propose Optimal ML iteration stopping method that determines the proper time to stop training ML model when this model reaches the minimum loss value. Our real experiments and large-scale simulation based on real trace show that MLFS reduces JCT by up to 53% and makespan by up to 52%, and improves accuracy by up to 64% when compared with existing ML job schedulers. We also open sourced our code.

Software-defined networking enabled big data tasks scheduling: A tabu search approach

The growth of information technology along with the revolution of the industry and business has led to the generation of an enormous amount of data. This big data needs a platform beyond the traditional data possessing context that relies on some computational servers communicating through a network in its lower layer. One of the most important challenges in data processing is how to transfer the big batches of data between the servers to achieve fast responsiveness. Consequently, the underlying network plays a critical role in the performance of a big data analysis platform. Ideally, this network must use the shortest path that has the lowest amount of load, to transfer the large-scale data. To address this issue, we propose a software-defined networking (SDN) enabled scheduling method that uses the tabu search algorithm to schedule big data tasks. The proposed algorithm not only considers data locality but also uses the network traffic status for efficient scheduling. Our extensive simulative study in the Mininet emulator shows that the proposed scheme gives high performance and minimizes job completion time.

Distributed job allocation using response threshold for heterogeneous robot team under deadline constraints

SummaryJob allocation among the robots is a significant challenge in a multi‐robot environment. This article proposes a distributed algorithm for job allocation with deadline constraints. It works utilizing the response threshold model inspired by job distribution among insect societies in nature. The team divides the jobs among themselves based on their threshold capacity. The proposed approach focus on maximizing the total jobs completed with the available resources. The static case where the robots never fail is considered first for evaluating the algorithm. Then it is run in a dynamic environment with robot failure to study the effects on job allocation. Two experimental studies are conducted with varying amounts of resources. In static and dynamic cases under experimental study I, the jobs completion percentage increased by an average of 37.6% and 35.8%, respectively, over the existing method. Under experimental study II, the proposed algorithm completed all the jobs with improved resource utilization. The results show that the proposed algorithm achieves better jobs completion. The algorithm is then tested in a real‐world experimental environment using FireBird V mobile robots under static and dynamic cases. The outcomes indicate that the proposed approach is efficient in job allocation in practical circumstances.

Design of deep learning model for radio resource allocation in 5G for massive iot device

With recent advancements in Internet of Things (IoT) communications, the demand for fastest communication is an utmost concern of the IoT technology. The advent of 5G telecommunication networks enables to bridge the demand on satisfying the Quality-of-Service (QoS) concerns in IoT communication. The massive number of devices in IoT communication is henceforth do not lie under limited resource allocation by embedding the 5G telecommunication network. The existing deep learning approaches lack in the allocation of resource in the energy constrained environment. In this paper, we address the above limitation of allocation the resource to prominent IoT devices using Deep Belief Network. It is efficient in the utilization of hidden layers and the probabilistic correlation among the instances are constructed effectively. In this Deep Belief Network, the adoption of rules will imply the resource allocation with the devices of top priority and it reduces based on the priority. DBN learns the context of the 5G and connected IoT devices whereas the resources are provisioned effectively through the learning. The proposed strategy beats the current state-of-the-art techniques, whereas the DBN achieves 74 % of successful job completion that is effective compared to other existing strategies. The simulation is performed to test the efficacy of Deep Learning on allocation 5G resources to IoT model. The results show that the proposed model achieves higher level of allocating the resources than other methods.

MPU-6050 Wheeled Robot Controlled Hand Gesture Using L298N Driver Based on Arduino

Robots are intelligent machines that are controlled by computer applications to perform various operations and services to complete high-risk jobs such as fire, poison, sharp objects, viruses, etc. MPU-6050 Serves as a movement medium in robot electronic systems where the robot moves. Arduino nano functions as a controller on a system robot that includes the MPU-6050, Motor Driver L298N, and NRF24L01. NRF24L01 Serves as a communication medium that is connected between the robot movement system and the hand gesture movement system. The L298N Motor Driver moves according to the signal sent by the MPU-6050 sensor. The L298N Motor Driver functions as an output medium that is made to the DC motor. The L298N Motor Driver will instruct the DC motor to move according to the signal sent by the MPU-6050 module. From the data obtained, the MPU-6050 works at a Y angle with an analog signal of 27.42 ° while driving. The MPU-6050 has X and Y axes with error values at X of 0.32% and Y of 0.48%. NRF24L01 functions as a command to send and receive data signals that work with a maximum range of 30 meters with a speed of 0.457 m/s.

On preemptive scheduling on unrelated machines using linear programming

&lt;abstract&gt;&lt;p&gt;We consider a basic preemptive scheduling problem where $ n $ non-simultaneously released jobs are to be processed by $ m $ unrelated parallel machines so as to minimize maximum job completion time. An optimal LP-solution has been used to construct an optimal preemptive schedule for simultaneously released jobs in time $ O(n^3) $. We propose fast $ O(m) $ time algorithm that finds an optimal schedule in case the above LP-solution possesses "small enough" number of non-zero elements. We propose another linear program for non-simultaneously released jobs and show how an optimal schedule can be constructed also in time $ O(m) $ from the optimal solution to that linear program. Based on another stronger linear program formulation, we extend the earlier known schedule construction procedure for non-simultaneously released jobs. The procedure is important, in particular, because there may exist no optimal schedule that agrees with an optimal LP-solution. An optimal LP-solution imposes a number of preemptions, and additional preemptions may occur during the schedule construction process, a job might be forced to be split on the same machine. We show that if no job split is allowed, even a restricted version of the problem on three unrelated machines is NP-hard. As a result, we obtain that, given an optimal LP-solution, it is NP-hard to find an optimal schedule that agrees with that LP-solution. As another side result, we obtain that it is NP-hard to find an optimal schedule with at most $ m-1 $ preemptions.&lt;/p&gt;&lt;/abstract&gt;