Cloud Scheduler: a resource manager for distributed compute clouds
P. Armstrong, A. Agarwal, A. Bishop, A. Charbonneau, R. Desmarais, K. Fransham, N. Hill, I. Gable, S. Gaudet, S. Goliath, R. Impey, C. Leavett-Brown, J. Ouellet… 展开作者
P. Armstrong, A. Agarwal, A. Bishop, A. Charbonneau, R. Desmarais, K. Fransham, N. Hill, I. Gable, S. Gaudet, S. Goliath, R. Impey, C. Leavett-Brown, J. Ouellete, M. Paterson, C. Pritchet, D. Penfold-Brown, W. Podaima, D. Schade, R. J. Sobie
The availability of Infrastructure-as-a-Service (IaaS) computing clouds gives researchers access to a large set of new resources for running complex scientific applications. However, exploiting cloud resources for large numbers of jobs requires significant effort and expertise. In order to make it simple and transparent for researchers to deploy their applications, we have developed a virtual machine resource manager (Cloud Scheduler) for distributed compute clouds. Cloud Scheduler boots and manages the user-customized virtual machines in response to a user's job submission. We describe the motivation and design of the Cloud Scheduler and present results on its use on both science and commercial clouds.
The performance of distributed and data-centric applications often critically depends on the interconnecting network. Emerging reconfigurable datacenter networks (RDCNs) are a particularly innovative approach to improve datacenter throughput. Relying on a dynamic optical topology which can be adjusted towards the workload in a demand-aware manner, RDCNs allow to exploit temporal and spatial locality in the communication pattern, and to provide topological shortcuts for frequently communicating racks. The key challenge, however, concerns how to realize demand-awareness in RDCNs in a scalable fa…
To solve the limitation of Hadoop on scalability, resource sharing, and application support, the open-source community proposes the next generation of Hadoop's compute platform called Yet Another Resource Negotiator (YARN) by separating resource management functions from the programming model. This separation enables various application types to run on YARN in parallel. To achieve fair resource sharing and high resource utilization, YARN provides the capacity scheduler and the fair scheduler. However, the performance impacts of the two schedulers are not clear when mixed applications run on a …
Task graphs provide a simple way to describe scientific workflows (sets of tasks with dependencies) that can be executed on both HPC clusters and in the cloud. An important aspect of executing such graphs is the used scheduling algorithm. Many scheduling heuristics have been proposed in existing works; nevertheless, they are often tested in oversimplified environments. We provide an extensible simulation environment designed for prototyping and benchmarking task schedulers, which contains implementations of various scheduling algorithms and is open-sourced, in order to be fully reproducible. W…
Parallel task-based programming models, like OpenMP, allow application developers to easily create a parallel version of their sequential codes. The standard OpenMP 4.0 introduced the possibility of describing a set of data dependences per task that the runtime uses to order the tasks execution. This order is calculated using shared graphs, which are updated by all threads in exclusive access using synchronization mechanisms (locks) to ensure the dependence management correctness. The contention in the access to these structures becomes critical in many-core systems because several threads may…
Large-scale distributed training is increasingly becoming communication bound. Many gradient compression algorithms have been proposed to reduce the communication overhead and improve scalability. However, it has been observed that in some cases gradient compression may even harm the performance of distributed training. In this paper, we propose MergeComp, a compression scheduler to optimize the scalability of communication-efficient distributed training. It automatically schedules the compression operations to optimize the performance of compression algorithms without the knowledge of model a…
In this paper, we introduce MARS, a new scheduling system for HPC-cloud infrastructures based on a cost-aware, flexible reinforcement learning approach, which serves as an intermediate layer for next generation HPC-cloud resource manager. MARS ensembles the pre-trained models from heuristic workloads and decides on the most cost-effective strategy for optimization. A whole workflow application would be split into several optimizable dependent sub-tasks, then based on the pre-defined resource management plan, a reward will be generated after executing a scheduled task. Lastly, MARS updates the …
Converged computing brings together the best of both worlds for high performance computing (HPC) and cloud-native communities. In fact, the economic impact of cloud-computing, and need for portability, flexibility, and manageability make it not important, but inevitable. Navigating this uncharted territory requires not just innovation in the technology space, but also effort toward collaboration and sharing of ideas. With these goals in mind, this work first tackles the central component of running batch workflows, whether in cloud or HPC: the workload manager. For cloud, Kubernetes has become…
This document describes the DIR net, a distributed environment which is part of the EFTOS fault tolerance framework. The DIR net is a system consisting of two components, called DIR Manager (or, shortly, the manager) and DIR Backup Agent (shortly, the backup). One manager and a set of backups is located in the system to be `guarded', one component per node. At this point the DIR net weaves a web which substantially does two things: 1) makes itself tolerant to a number of possible faults, and 2) gathers information pertaining the run of the user application. As soon as an error occurs within th…
Cloud Computing is emerging as a new computational paradigm shift. Hadoop-MapReduce has become a powerful Computation Model for processing large data on distributed commodity hardware clusters such as Clouds. In all Hadoop implementations, the default FIFO scheduler is available where jobs are scheduled in FIFO order with support for other priority based schedulers also. In this paper we study various scheduler improvements possible with Hadoop and also provided some guidelines on how to improve the scheduling in Hadoop in Cloud Environments.
As grids are in essence heterogeneous, dynamic, shared and distributed environments, managing these kinds of platforms efficiently is extremely complex. A promising scalable approach to deal with these intricacies is the design of self-managing of autonomic applications. Autonomic applications adapt their execution accordingly by considering knowledge about their own behaviour and environmental conditions.QoS based User Driven scheduling for grid that provides the self-optimizing ability in autonomic applications. Computational grids to provide a user to solve large scale problem by spreading …
Large-scale interactive web services and advanced AI applications make sophisticated decisions in real-time, based on executing a massive amount of computation tasks on thousands of servers. Task schedulers, which often operate in heterogeneous and volatile environments, require high throughput, i.e., scheduling millions of tasks per second, and low latency, i.e., incurring minimal scheduling delays for millisecond-level tasks. Scheduling is further complicated by other users' workloads in a shared system, other background activities, and the diverse hardware configurations inside datacenters.…
This paper discusses the deployment of existing render farm manager in a typical compute cluster environment such as a university. Usually, both a render farm and a compute cluster use different queue managers and assume total control over the physical resources. But, taking out the physical resources from an existing compute cluster in a university-like environment whose primary use of the cluster is to run numerical simulations may not be possible. It can potentially reduce the overall resource utilization in a situation where compute tasks are more than rendering tasks. Moreover, it can inc…
Distributed data processing systems like MapReduce, Spark, and Flink are popular tools for analysis of large datasets with cluster resources. Yet, users often overprovision resources for their data processing jobs, while the resource usage of these jobs also typically fluctuates considerably. Therefore, multiple jobs usually get scheduled onto the same shared resources to increase the resource utilization and throughput of clusters. However, job runtimes and the utilization of shared resources can vary significantly depending on the specific combinations of co-located jobs. This paper presents…
Whether it is for audit or for recovery purposes, data checkpointing is an important problem of distributed database systems. Actually, transactions establish dependence relations on data checkpoints taken by data object managers. So, given an arbitrary set of data checkpoints (including at least a single data checkpoint from a data manager, and at most a data checkpoint from each data manager), an important question is the following one: ``Can these data checkpoints be members of a same consistent global checkpoint?''. This paper answers this question by providing a necessary and sufficient c…
This paper introduces a resource allocation framework specifically tailored for addressing the problem of dynamic placement (or pinning) of parallelized applications to processing units. Under the proposed setup each thread of the parallelized application constitutes an independent decision maker (or agent), which (based on its own prior performance measurements and its own prior CPU-affinities) decides on which processing unit to run next. Decisions are updated recursively for each thread by a resource manager/scheduler which runs in parallel to the application's threads and periodically reco…
Applications in science and engineering often require huge computational resources for solving problems within a reasonable time frame. Parallel supercomputers provide the computational infrastructure for solving such problems. A traditional application scheduler running on a parallel cluster only supports static scheduling where the number of processors allocated to an application remains fixed throughout the lifetime of execution of the job. Due to the unpredictability in job arrival times and varying resource requirements, static scheduling can result in idle system resources thereby decrea…
Scientific workflow management systems like Nextflow support large-scale data analysis by abstracting away the details of scientific workflows. In these systems, workflows consist of several abstract tasks, of which instances are run in parallel and transform input partitions into output partitions. Resource managers like Kubernetes execute such workflow tasks on cluster infrastructures. However, these resource managers only consider the number of CPUs and the amount of available memory when assigning tasks to resources; they do not consider hardware differences beyond these numbers, while com…
Efficient scheduling of distributed deep learning (DL) jobs in large GPU clusters is crucial for resource efficiency and job performance. While server sharing among jobs improves resource utilization, interference among co-located DL jobs occurs due to resource contention. Interference-aware job placement has been studied, with white-box approaches based on explicit interference modeling and black-box schedulers with reinforcement learning. In today's clusters containing thousands of GPU servers, running a single scheduler to manage all arrival jobs in a timely and effective manner is challeng…
Apache Mesos, a two-level resource scheduler, provides resource sharing across multiple users in a multi-tenant cluster environment. Computational resources (i.e., CPU, memory, disk, etc. ) are distributed according to the Dominant Resource Fairness (DRF) policy. Mesos frameworks (users) receive resources based on their current usage and are responsible for scheduling their tasks within the allocation. We have observed that multiple frameworks can cause fairness imbalance in a multiuser environment. For example, a greedy framework consuming more than its fair share of resources can deny resour…
The use of meta-schedulers for resource management in large-scale distributed systems often leads to a hierarchy of schedulers. In this paper, we discuss why existing meta-scheduling hierarchies are sometimes not sufficient for Grid systems due to their inability to re-organise jobs already scheduled locally. Such a job re-organisation is required to adapt to evolving loads which are common in heavily used Grid infrastructures. We propose a peer-to-peer scheduling model and evaluate it using case studies and mathematical modelling. We detail the DIANA (Data Intensive and Network Aware) schedul…
Cloud computing has been developed as one of the prominent paradigm for providing on demand resources to the end user based on signed service level agreement and pay as use model. Cloud computing provides resources using multitenant architecture where infrastructure is generated from multiple or single geographical distributed cloud datacenters. Scheduling of cloud application requests to cloud infrastructure is one of the main research area in cloud computing. Researchers have developed many scheduling applications for which they have used different simulators available in the market such as …
Cloud resource management is often modeled by two-dimensional bin packing with a set of items that correspond to tasks having fixed CPU and memory requirements. However, applications running in clouds are much more flexible: modern frameworks allow to (horizontally) scale a single application to dozens, even hundreds of instances; and then the load balancer can precisely divide the workload between them. We analyze a model that captures this (semi)-flexibility of cloud resource management. Each cloud application is characterized by its memory footprint and its momentary CPU load. Combining the…
In this paper, we address distributed convergence to fair allocations of CPU resources for time-sensitive applications. We propose a novel resource management framework where a centralized objective for fair allocations is decomposed into a pair of performance-driven recursive processes for updating: (a) the allocation of computing bandwidth to the applications (resource adaptation), executed by the resource manager, and (b) the service level of each application (service-level adaptation), executed by each application independently. We provide conditions under which the distributed recursive s…
Application service providers (ASPs) now develop, deploy, and maintain complex computing platforms within multiple cloud infrastructures to improve resilience, responsiveness and elasticity of their applications. On the other hand, complex applications have little control and visibility over network resources, and need to use low-level hacks to extract network properties and prioritize traffic. This biased view, limits tenants flexibility while deploying their applications and prevents them from implementing part of the application logic in the network. In this paper, we propose the CNSMO (CYC…
Wider adoption of the Grid concept has led to an increasing amount of federated computational, storage and visualisation resources being available to scientists and researchers. Distributed and heterogeneous nature of these resources renders most of the legacy cluster monitoring and management approaches inappropriate, and poses new challenges in workflow scheduling on such systems. Effective resource utilisation monitoring and highly granular yet adaptive measurements are prerequisites for a more efficient Grid scheduler. We present a suite of measurement applications able to monitor per-proc…
In this paper, a method for efficient scheduling to obtain optimum job throughput in a distributed campus grid environment is presented; Traditional job schedulers determine job scheduling using user and job resource attributes. User attributes are related to current usage, historical usage, user priority and project access. Job resource attributes mainly comprise of soft requirements (compilers, libraries) and hard requirements like memory, storage and interconnect. A job scheduler dispatches jobs to a resource if a job's hard and soft requirements are met by a resource. In current scenario d…
Whilst computational resources at the cloud edge can be leveraged to improve latency and reduce the costs of cloud services for a wide variety mobile, web, and IoT applications; such resources are naturally constrained. For distributed stream processing applications, there are clear advantages to offloading some processing work to the cloud edge. Many state of the art stream processing applications such as Flink and Spark Streaming, being designed to run exclusively in the cloud, are a poor fit for such hybrid edge/cloud deployment settings, not least because their schedulers take limited cons…
Soteria is a user right management system designed to safeguard user-data privacy in a transparent and provable manner in compliance to regulations such as GDPR and CCPA. Soteria represents user data rights as formal executable sharing agreements, which can automatically be translated into a human readable form and enforced as data are queried. To support revocation and to prove compliance, an indelible, audited trail of the hash of data access and sharing agreements are stored on a two-layer distributed ledger. The main chain ensures partition tolerance and availability (PA) properties while …
Clusters, grids, and peer-to-peer (P2P) networks have emerged as popular paradigms for next generation parallel and distributed computing. The management of resources and scheduling of applications in such large-scale distributed systems is a complex undertaking. In order to prove the effectiveness of resource brokers and associated scheduling algorithms, their performance needs to be evaluated under different scenarios such as varying number of resources and users with different requirements. In a grid environment, it is hard and even impossible to perform scheduler performance evaluation in …
We show that distributed Infrastructure-as-a-Service (IaaS) compute clouds can be effectively used for the analysis of high energy physics data. We have designed a distributed cloud system that works with any application using large input data sets requiring a high throughput computing environment. The system uses IaaS-enabled science and commercial clusters in Canada and the United States. We describe the process in which a user prepares an analysis virtual machine (VM) and submits batch jobs to a central scheduler. The system boots the user-specific VM on one of the IaaS clouds, runs the job…
Research interest in Grid computing has grown significantly over the past five years. Management of distributed resources is one of the key issues in Grid computing. Central to management of resources is the effectiveness of resource allocation as it determines the overall utility of the system. The current approaches to superscheduling in a grid environment are non-coordinated since application level schedulers or brokers make scheduling decisions independently of the others in the system. Clearly, this can exacerbate the load sharing and utilization problems of distributed resources due to s…
Clusters of computers have emerged as mainstream parallel and distributed platforms for high-performance, high-throughput and high-availability computing. To enable effective resource management on clusters, numerous cluster managements systems and schedulers have been designed. However, their focus has essentially been on maximizing CPU performance, but not on improving the value of utility delivered to the user and quality of services. This paper presents a new computational economy driven scheduling system called Libra, which has been designed to support allocation of resources based on the…
A distributed heap storage manager has been implemented on the Fujitsu AP1000 multicomputer. The performance of various pre-fetching strategies is experimentally compared. Subjective programming benefits and objective performance benefits of up to 10% in pre-fetching are found for certain applications, but not for all. The performance benefits of pre-fetching depend on the specific data structure and access patterns. We suggest that control of pre-fetching strategy be dynamically under the control of the application.
Training and deploying deep learning models in real-world applications require processing large amounts of data. This is a challenging task when the amount of data grows to a hundred terabytes, or even, petabyte-scale. We introduce a hybrid distributed cloud framework with a unified view to multiple clouds and an on-premise infrastructure for processing tasks using both CPU and GPU compute instances at scale. The system implements a distributed file system and failure-tolerant task processing scheduler, independent of the language and Deep Learning framework used. It allows to utilize unstable…
CoRaiS: Lightweight Real-Time Scheduler for Multi-Edge Cooperative Computing
Purple Mountain Laboratories（紫金山实验室）；School of Computer Science, Northwestern Polytechnical University（西北工业大学计算机学院）；National University of Defense Technology（国防科技大学）；State Key Laboratory of networking and Switching Technology, Beijing University of Posts and Telecommunications（北京邮电大学）；Carleton University（卡尔顿大学）
Data centers are giant factories of Internet data and services. Worldwide data centers consume energy and emit emissions more than airline industry. Unfortunately, most of data centers are significantly underutilized. One of the major reasons is the big gaps between the real usage and the provisioned resources because users tend to over-estimate their demand and data center operators often rely on users' requests for resource allocation. In this paper, we first conduct an in-depth analysis of a Google cluster trace to unveil the root causes for low utilization and highlight the great potential…
In this paper, the resource management problem in geographically distributed cloud systems is considered. The Follow Me Cloud concept which enables service migration across federated data centers (DCs) is adopted. Therefore, there are two types of service requests to the DC, i.e., new requests (NRs) initiated in the local service area and migration requests (MRs) generated when mobile users move across service areas. A novel resource management scheme is proposed to help the resource manager decide whether to accept the service requests (NRs or MRs) or not and determine how much resources shou…
Deep Reinforcement Learning (DRL) is being increasingly applied to the problem of resource allocation for emerging System-on-Chip (SoC) applications, and has shown remarkable promises. In this paper, we introduce SoCRATES (SoC Resource AdapTivE Scheduler), an extremely efficient DRL-based SoC scheduler which maps a wide range of hierarchical jobs to heterogeneous resources within SoC using the Eclectic Interaction Matching (EIM) technique. It is noted that the majority of SoC resource management approaches have been targeting makespan minimization with fixed number of jobs in the system. In co…
Scientists increasingly rely on Python tools to perform scalable distributed memory array operations using rich, NumPy-like expressions. However, many of these tools rely on dynamic schedulers optimized for abstract task graphs, which often encounter memory and network bandwidth-related bottlenecks due to sub-optimal data and operator placement decisions. Tools built on the message passing interface (MPI), such as ScaLAPACK and SLATE, have better scaling properties, but these solutions require specialized knowledge to use. In this work, we present NumS, an array programming library which optim…
To reduce uploading bandwidth and address privacy concerns, deep learning at the network edge has been an emerging topic. Typically, edge devices collaboratively train a shared model using real-time generated data through the Parameter Server framework. Although all the edge devices can share the computing workloads, the distributed training processes over edge networks are still time-consuming due to the parameters and gradients transmission procedures between parameter servers and edge devices. Focusing on accelerating distributed Convolutional Neural Networks (CNNs) training at the network …
"Science gateway" (SG) ideology means a user-friendly intuitive interface between scientists (or scientific communities) and different software components + various distributed computing infrastructures (DCIs) (like grids, clouds, clusters), where researchers can focus on their scientific goals and less on peculiarities of software/DCI. "IMP Science Gateway Portal" (http://scigate.imp.kiev.ua) for complex workflow management and integration of distributed computing resources (like clusters, service grids, desktop grids, clouds) is presented. It is created on the basis of WS-PGRADE and gUSE tec…
