Multiple Concurrent Jobs Research Articles

Differentiate Quality of Experience Scheduling for Deep Learning Inferences With Docker Containers in the Cloud

With the prevalence of big-data-driven applications, such as face recognition on smartphones and tailored recommendations from Google Ads, we are on the road to a lifestyle with significantly more intelligence than ever before. Various neural network powered models are running at the back end of their intelligence to enable quick responses to users. Supporting those models requires lots of cloud-based computational resources, e.g., CPUs and GPUs. The cloud providers charge their clients by the amount of resources that they occupy. Clients have to balance the budget and quality of experiences (e.g., response time). The budget leans on individual business owners, and the required Quality of Experience (QoE) depends on usage scenarios of different applications. For instance, an autonomous vehicle requires an real-time response, but unlocking your smartphone can tolerate delays. However, cloud providers fail to offer a QoE-based option to their clients. In this paper, we propose DQoES, differentiated quality of experience scheduler for deep learning inferences. DQoES accepts clients' specifications on targeted QoEs, and dynamically adjusts resources to approach their targets. Through the extensive cloud-based experiments, DQoES demonstrates that it can schedule multiple concurrent jobs with respect to various QoEs and achieve up to 8x times more satisfied models when compared to the existing system.

PAS: Performance-Aware Job Scheduling for Big Data Processing Systems

Big data analytics has become increasingly vital in many modern enterprise applications such as user profiling and business process optimization. Today’s big data processing systems, such as Hadoop MapReduce, Spark, and Hive, treat big data applications as a batch of jobs for scheduling. Existing schedulers in production systems often maintain fair allocation without considering application performance and resource utilization simultaneously. It is challenging to perform job scheduling in big data systems to achieve both low turnaround time and high resource utilization due to the high complexity in data processing logics and the dynamic variation in workloads. In this article, we propose a performance-aware scheduler, referred to as PAS, which dynamically schedules big data jobs in Hadoop YARN and Spark and autonomously adjusts scheduling policies to improve application performance and resource utilization. Specifically, PAS schedules multiple concurrent jobs using different policies based on the predicted job completion time and employs a greedy approach and a one-step lookahead strategy to opportunistically maximize the average job performance while still maintaining a satisfactory level of resource utilization. We implement PAS in Hadoop YARN and evaluate its performance with HiBench, a well-known big data processing benchmark. Experimental results show that PAS reduces the average turnaround time by 25% and the makespan by 15% in comparison with four state-of-the-art schedulers.

A Path Relinking Method for the Joint Online Scheduling and Capacity Allocation of DL Training Workloads in GPU as a Service Systems

The Deep Learning (DL) paradigm gained remarkable popularity in recent years. DL models are used to tackle increasingly complex problems, making the training process require considerable computational power. The parallel computing capabilities offered by modern GPUs partially fulfill this need, but the high costs related to GPU as a Service solutions in the cloud call for efficient capacity planning and job scheduling algorithms to reduce operational costs via resource sharing. In this work, we jointly address the online capacity planning and job scheduling problems from the perspective of cloud end-users. We present a Mixed Integer Linear Programming (MILP) formulation, and a path relinking-based method aiming at optimizing operational costs by (i) rightsizing Virtual Machine (VM) capacity at each node, (ii) partitioning the set of GPUs among multiple concurrent jobs on the same VM, and (iii) determining a due-date-aware job schedule. An extensive experimental campaign attests the effectiveness of the proposed approach in practical scenarios: costs savings up to 97% are attained compared with first-principle methods based on, e.g., Earliest Deadline First, cost reductions up to 20% are obtained with respect to a previously proposed Hierarchical Method and up to 95% against a dynamic programming-based method from the literature. Scalability analyses show that systems with up to 100 nodes and 450 concurrent jobs can be managed in less than 7 seconds. The validation in a prototype cloud environment shows a deviation below 5% between real and predicted costs.

Twister2 Cross‐platform resource scheduler for big data

AbstractTwister2 is an open‐source big data hosting environment designed to process both batch and streaming data at scale. Twister2 runs jobs in both high‐performance computing (HPC) and big data clusters. It provides a cross‐platform resource scheduler to run jobs in diverse environments. Twister2 is designed with a layered architecture to support various clusters and big data problems. In this paper, we present the cross‐platform resource scheduler of Twister2. We identify required services and explain implementation details. We present job startup delays for single jobs and multiple concurrent jobs in Kubernetes and OpenMPI clusters. We compare job startup delays for Twister2 and Spark at a Kubernetes cluster. In addition, we compare the performance of terasort algorithm on Kubernetes and bare metal clusters at AWS cloud.

Health workforce perspectives of barriers inhibiting the provision of quality care in Nepal and Somalia – A qualitative study

ObjectiveIn this paper settings from Nepal and Somalia are used to focus on the perspectives of healthcare providers within two fragile health systems. The objective of this study was to describe barriers inhibiting quality healthcare in Nepal and Somalia from a health workforce perspective. MethodsData were collected through 19 semi-structured interviews with healthcare providers working in healthcare facilities. Ten interviews were conducted in Nepal and nine in Somalia. ResultsVarious structural barriers inhibiting the availability, accessibility, and acceptability of the quality care were similar in both countries. Barriers inhibiting the availability of quality care were linked to healthcare providers being overburdened with multiple concurrent jobs. Barriers inhibiting the accessibility to quality healthcare included long distances and the uncertain availability of transportation, and barriers to acceptability of quality healthcare was inhibited by a lack of respect from healthcare providers, characterised by neglect, verbal abuse, and lack of competence. ConclusionsInequality, poverty, traditional and cultural practices plus the heavy burden placed on healthcare providers are described as the underlying causes of the poor provision of quality care and the consequential shortcomings that emerge from it. In order to improve this situation adequate planning and policies that support the deployment and retention of the healthcare providers and its equitable distribution is required. Another important aspect is provision of training to equip healthcare providers with the ability to provide respectful quality care in order for the population to enjoy good standard of healthcare services.

Processing Concurrent Graph Analytics with Decoupled Computation Model

Graph processing systems have been widely used in enterprises like online social networks to process their daily jobs. With the fast growing of social applications, they have to efficiently handle massive concurrent jobs. However, due to the inherent design for a single job, existing systems incur great inefficiency in terms of memory usage, execution and fault tolerance. Motivated by this issue, in this paper we introduce Seraph, a graph processing system that enables efficient job-level parallelism. Seraph is designed based on a decoupled computation model, which decouples both the runtime job data and the computation logic. Decoupling the runtime data allows multiple concurrent jobs to share graph structure data in memory, which fundamentally increases job-level concurrency and reduces fault tolerance overhead. Decoupling computation logic could extend scheduling space, which benefits both execution performance and memory consumption. Seraph adopts a copy-on-write semantic to isolate the graph mutation of concurrent jobs, and a lazy snapshot protocol to generate consistent graph snapshots for jobs submitted at different time. Based on the decoupled model, it provides unified programming interfaces for both synchronous and asynchronous graph applications. Moreover, Seraph implements a lightweight checkpoint mechanism which can tremendously reduce the fault tolerance overhead. The evaluation results show that Seraph significantly outperforms popular systems (such as Giraph, Spark, GraphX and PowerLyra) in both memory usage and job completion time, when executing concurrent graph jobs.

A Comprehensive Study of MapReduce Over Lustre for Intermediate Data Placement and Shuffle Strategies on HPC Clusters

With high performance interconnects and parallel file systems, running MapReduce over modern High Performance Computing (HPC) clusters has attracted much attention due to its uniqueness of solving data analytics problems with a combination of Big Data and HPC technologies. Since the MapReduce architecture relies heavily on the availability of local storage media, the Lustre-based global storage in HPC clusters poses many new opportunities and challenges. In this paper, we perform a comprehensive study on different MapReduce over Lustre deployments and propose a novel high-performance design of YARN MapReduce on HPC clusters by utilizing Lustre as the additional storage provider for intermediate data. With a deployment architecture where both local disks and Lustre are utilized for intermediate data storage, we propose a novel priority directory selection scheme through which RDMA-enhanced MapReduce can choose the best intermediate storage during runtime by on-line profiling. Our results indicate that, we can achieve 44 percent performance benefit for shuffle-intensive workloads in leadership-class HPC systems. Our priority directory selection scheme can improve the job execution time by 63 percent over default MapReduce while executing multiple concurrent jobs. To the best of our knowledge, this is the first such comprehensive study for YARN MapReduce with Lustre and RDMA.

On the optimization of schedules for MapReduce workloads in the presence of shared scans

We consider MapReduce clusters designed to support multiple concurrent jobs, concentrating on environments in which the number of distinct datasets is modest relative to the number of jobs. In such scenarios, many individual datasets are likely to be scanned concurrently by multiple Map phase jobs. As has been noticed previously, this scenario provides an opportunity for Map phase jobs to cooperate, sharing the scans of these datasets, and thus reducing the costs of such scans. Our paper has three main contributions over previous work. First, we present a novel and highly general method for sharing scans and thus amortizing their costs. This concept, which we call cyclic piggybacking, has a number of advantages over the more traditional batching scheme described in the literature. Second, we notice that the various subjobs generated in this manner can be assumed in an optimal schedule to respect a natural chain precedence ordering. Third, we describe a significant but natural generalization of the recently introduced FLEX scheduler for optimizing schedules within the context of this cyclic piggybacking paradigm, which can be tailored to a variety of cost metrics. Such cost metrics include average response time, average stretch, and any minimax-type metric--a total of 11 separate and standard metrics in all. Moreover, most of this carries over in the more general case of overlapping rather than identical datasets as well, employing what we will call semi-shared scans. In such scenarios, chain precedence is replaced by arbitrary precedence, but we can still handle 8 of the original 11 metrics. The overall approach, including both cyclic piggybacking and the FLEX scheduling generalization, is called CIRCUMFLEX. We describe some practical implementation strategies. And we evaluate the performance of CIRCUMFLEX via a variety of simulation and real benchmark experiments.

CIRCUMFLEX

We consider MapReduce clusters designed to support multiple concurrent jobs, concentrating on environments in which the number of distinct datasets is modest relative to the number of jobs. Many datasets in such scenarios wind up being scanned by multiple concurrent Map phase jobs. As has been noticed previously, this scenario provides an opportunity for Map phase jobs to cooperate , sharing the scans of these datasets, and thus reducing the costs of such scans. Our paper has two main contributions. First, we present a novel and highly general method for sharing scans and thus amortizing their costs. This concept, which we call cyclic piggybacking , has a number of advantages over the more traditional batching scheme described in the literature. Second, we describe a significant but natural generalization of the recently introduced flex scheduler, for optimizing schedules within the context of this cyclic piggybacking paradigm. The overall approach, including both cyclic piggybacking and the flex generalization, is called circumflex. We demonstrate the excellent performance of circumflex via a variety of simulation experiments.

Efficient resource management for running multiple concurrent jobs in a computational grid environment

Managing resources in Grid is an intricate issue and is not possible at the application or user level. Grid provides uniform access to heterogeneous resources owned by multiple organizations. Any time a new resource can join the Grid or an existing resource can withdraw. The system loads and status of the resources also change frequently. Thus, a resource broker is an essential component in a Grid environment that can assist in the selection of a right resource provider for a job in all aspects. This paper discusses the design and implementation of resource brokering strategies within the multi-agent framework. These strategies help in finding out an optimal allocation of resources for executing multiple concurrent jobs in a Grid environment. Different stages in resource brokering and their implementation within the framework are discussed. The paper also presents results of evaluating the strategies and demonstrates their effectiveness.