Distributed Deep Learning (DDL) has rapidly grown its popularity since it helps boost the training performance on high-performance GPU clusters. Efficient job scheduling is indispensable to maximize the overall performance of the cluster when training multiple jobs simultaneously. However, existing schedulers do not consider the communication contention of multiple communication tasks from different distributed training jobs, which could deteriorate the system performance and prolong the job completion time. In this paper, we first establish a new DDL job scheduling framework which organizes DDL jobs as Directed Acyclic Graphs (DAGs) and considers communication contention between nodes. We then propose an efficient algorithm, LWF-$κ$, to balance the GPU utilization and consolidate the allocated GPUs for each job. When scheduling those communication tasks, we observe that neither avoiding all the contention nor blindly accepting them is optimal to minimize the job completion time. We thus propose a provable algorithm, AdaDUAL, to efficiently schedule those communication tasks. Based on AdaDUAL, we finally propose Ada-SRSF for the DDL job scheduling problem. Simulations on a 64-GPU cluster connected with 10 Gbps Ethernet show that LWF-$κ$ achieves up to $1.59\times$ improvement over the classical first-fit algorithms. More importantly, Ada-SRSF reduces the average job completion time by $20.1\%$ and $36.7\%$, as compared to the SRSF(1) scheme (avoiding all the contention) and the SRSF(2) scheme (blindly accepting all of two-way communication contention) respectively.
Powered by advances in deep learning (DL) techniques, machine learning and artificial intelligence have achieved astonishing successes. However, the rapidly growing needs for DL also led to communication- and resource-intensive distributed training jobs for large-scale DL training, which are typically deployed over GPU clusters. To sustain the ever-increasing demand for DL training, the so-called "ring-all-reduce" (RAR) technologies have recently emerged as a favorable computing architecture to efficiently process network communication and computation load in GPU clusters. The most salient fea…
The University of Tokyo（东京大学）；Preferred Networks, Inc.（优选网络公司）
One significant challenge in the job scheduling of computing clusters for the development of deep learning algorithms is the efficient scheduling of trial-and-error (TE) job, the type of job in which the users seek to conduct small-scale experiments while monitoring their processes. Unfortunately, the existing job schedulers to date do not feature well-balanced scheduling for the mixture of TE jobs and best-effort (BE) jobs, or they can handle the mixture in limited situations at most. To fill in this niche, we propose an algorithm that can significantly reduce the latency of TE jobs in versat…
Diverse workloads such as interactive supercomputing, big data analysis, and large-scale AI algorithm development, requires a high-performance scheduler. This paper presents a novel node-based scheduling approach for large scale simulations of short running jobs on MIT SuperCloud systems, that allows the resources to be fully utilized for both long running batch jobs while simultaneously providing fast launch and release of large-scale short running jobs. The node-based scheduling approach has demonstrated up to 100 times faster scheduler performance that other state-of-the-art systems.
Coflow is a recently proposed networking abstraction to help improve the communication performance of data-parallel computing jobs. In multi-stage jobs, each job consists of multiple coflows and is represented by a Directed Acyclic Graph (DAG). Efficiently scheduling coflows is critical to improve the data-parallel computing performance in data centers. Compared with hand-tuned scheduling heuristics, existing work DeepWeave [1] utilizes Reinforcement Learning (RL) framework to generate highly-efficient coflow scheduling policies automatically. It employs a graph neural network (GNN) to encode …
MapReduce is emerged as a prominent programming model for data-intensive computation. In this work, we study power-aware MapReduce scheduling in the speed scaling setting first introduced by Yao et al. [FOCS 1995]. We focus on the minimization of the total weighted completion time of a set of MapReduce jobs under a given budget of energy. Using a linear programming relaxation of our problem, we derive a polynomial time constant-factor approximation algorithm. We also propose a convex programming formulation that we combine with standard list scheduling policies, and we evaluate their performan…
Energy consumption is one of the most critical concerns in designing computing devices, ranging from portable embedded systems to computer cluster systems. Furthermore, in the past decade, cluster systems have increasingly risen as popular platforms to run computing-intensive real-time applications in which the performance is of great importance. However, due to different characteristics of real-time workloads, developing general job scheduling solutions that efficiently address both energy consumption and performance in real-time cluster systems is a challenging problem. In this paper, inspir…
Baidu Inc.（百度公司）；Soochow University（苏州大学）；Auburn University（奥本大学）；North Carolina State University（北卡罗来纳州立大学）
Recent years have witnessed a large amount of decentralized data in various (edge) devices of end-users, while the decentralized data aggregation remains complicated for machine learning jobs because of regulations and laws. As a practical approach to handling decentralized data, Federated Learning (FL) enables collaborative global machine learning model training without sharing sensitive raw data. The servers schedule devices to jobs within the training process of FL. In contrast, device scheduling with multiple jobs in FL remains a critical and open problem. In this paper, we propose a novel…
Training machine learning (ML) models on large datasets requires considerable computing power. To speed up training, it is typical to distribute training across several machines, often with specialized hardware like GPUs or TPUs. Managing a distributed training job is complex and requires dealing with resource contention, distributed configurations, monitoring, and fault tolerance. In this paper, we describe TonY, an open-source orchestrator for distributed ML jobs built at LinkedIn to address these challenges.
We consider a distributed computing network consisting of a master and multiple workers processing tasks of different types. The master is running multiple applications. Each application stochastically generates real-time jobs with a strict job deadline, where each job is a collection of tasks of some types specified by the application. A real-time job is completed only when all its tasks are completed by the corresponding workers within the deadline. Moreover, we consider unreliable workers, whose processing speeds are uncertain. Because of the limited processing abilities of the workers, an …
In this paper, we study the market-oriented online bi-objective service scheduling problem for pleasingly parallel jobs with variable resources in cloud environments, from the perspective of SaaS (Software-as-as-Service) providers who provide job-execution services. The main process of scheduling SaaS services in clouds is: a SaaS provider purchases cloud instances from IaaS providers to schedule end users' jobs and charges users accordingly. This problem has several particular features, such as the job-oriented end users, the pleasingly parallel jobs with soft deadline constraints, the online…
We study the performance of non-adaptive scheduling policies in computing systems with multiple servers. Compute jobs are mostly regular, with modest service requirements. However, there are sporadic data intensive jobs, whose expected service time is much higher than that of the regular jobs. Forthis model, we are interested in the effect of scheduling policieson the average time a job spends in the system. To this end, we introduce two performance indicators in a simplified, only-arrival system. We believe that these performance indicators are good predictors of the relative performance of t…
There is a general trend towards solving problems suited to deep learning with more complex deep learning architectures trained on larger training sets. This requires longer compute times and greater data parallelization or model parallelization. Both data and model parallelism have been historically faster in parameter server architectures, but data parallelism is starting to be faster in ring architectures due to algorithmic improvements. In this paper, we analyze the math behind ring architectures and make an informed adaptation of dynamic scheduling to ring architectures. To do so, we form…
Database platform-as-a-service (dbPaaS) is developing rapidly and a large number of databases have been migrated to run on the Clouds for the low cost and flexibility. Emerging Clouds rely on the tenants to provide the resource specification for their database workloads. However, they tend to over-estimate the resource requirement of their databases, resulting in the unnecessarily high cost and low Cloud utilization. A methodology that automatically suggests the "just-enough" resource specification that fulfills the performance requirement of every database workload is profitable. To this end,…
The proliferation of multi-core and multiprocessor-based computer systems has led to explosive development of parallel applications and hence the need for efficient schedulers. In this paper, we study hierarchical scheduling for malleable parallel jobs on multiprocessor-based systems, which appears in many distributed and multilayered computing environments. We propose a hierarchical scheduling algorithm, named AC-DS, that consists of a feedback-driven adaptive scheduler, a desire aggregation scheme and an efficient resource allocation policy. From theoretical perspective, we show that AC-DS h…
Advance reservation is important to guarantee the quality of services of jobs by allowing exclusive access to resources over a defined time interval on resources. It is a challenge for the scheduler to organize available resources efficiently and to allocate them for parallel AR jobs with deadline constraint appropriately. This paper provides a slot-based data structure to organize available resources of multiprocessor systems in a way that enables efficient search and update operations, and formulates a suite of scheduling policies to allocate resources for dynamically arriving AR requests. T…
Fueled by advances in distributed deep learning (DDL), recent years have witnessed a rapidly growing demand for resource-intensive distributed/parallel computing to process DDL computing jobs. To resolve network communication bottleneck and load balancing issues in distributed computing, the so-called ``ring-all-reduce'' decentralized architecture has been increasingly adopted to remove the need for dedicated parameter servers. To date, however, there remains a lack of theoretical understanding on how to design resource optimization algorithms for efficiently scheduling ring-all-reduce DDL job…
The scheduling literature has traditionally focused on a single type of resource (e.g., computing nodes). However, scientific applications in modern High-Performance Computing (HPC) systems process large amounts of data, hence have diverse requirements on different types of resources (e.g., cores, cache, memory, I/O). All of these resources could potentially be exploited by the runtime scheduler to improve the application performance. In this paper, we study multi-resource scheduling to minimize the makespan of computational workflows comprised of parallel jobs subject to precedence constraint…
In job scheduling, the concept of malleability has been explored since many years ago. Research shows that malleability improves system performance, but its utilization in HPC never became widespread. The causes are the difficulty in developing malleable applications, and the lack of support and integration of the different layers of the HPC software stack. However, in the last years, malleability in job scheduling is becoming more critical because of the increasing complexity of hardware and workloads. In this context, using nodes in an exclusive mode is not always the most efficient solution…
We describe in this paper a new method for building an efficient algorithm for scheduling jobs in a cluster. Jobs are considered as parallel tasks (PT) which can be scheduled on any number of processors. The main feature is to consider two criteria that are optimized together. These criteria are the makespan and the weighted minimal average completion time (minsum). They are chosen for their complementarity, to be able to represent both user-oriented objectives and system administrator objectives. We propose an algorithm based on a batch policy with increasing batch sizes, with a smart selecti…
We consider a natural scheduling problem which arises in many distributed computing frameworks. Jobs with diverse resource requirements (e.g. memory requirements) arrive over time and must be served by a cluster of servers, each with a finite resource capacity. To improve throughput and delay, the scheduler can pack as many jobs as possible in the servers subject to their capacity constraints. Motivated by the ever-increasing complexity of workloads in shared clusters, we consider a setting where the jobs' resource requirements belong to a very large number of diverse types or, in the extreme,…
This paper provides three nearly-optimal algorithms for scheduling $t$ jobs in the $\mathsf{CLIQUE}$ model. First, we present a deterministic scheduling algorithm that runs in $O(\mathsf{GlobalCongestion} + \mathsf{dilation})$ rounds for jobs that are sufficiently efficient in terms of their memory. The $\mathsf{dilation}$ is the maximum round complexity of any of the given jobs, and the $\mathsf{GlobalCongestion}$ is the total number of messages in all jobs divided by the per-round bandwidth of $n^2$ of the $\mathsf{CLIQUE}$ model. Both are inherent lower bounds for any scheduling algorithm. …
This paper presents a multiagent approach as a paradigm for scheduling parallel jobs in a parallel system. Scheduling parallel jobs is performed as a means to balance the load of a system in order to improve the performance of a parallel application. Parallel job scheduling is presented as a mapping between two graphs: one represents the dependency of jobs and the other represents the interconnection among processors. The usual implementation of parallel job scheduling algorithms is via the master-slave paradigm. The master-slave paradigm has inherent communication bottleneck that reduces the …
We study the problem of scheduling jobs on fault-prone machines communicating via a shared channel, also known as multiple-access channel. We have $n$ arbitrary length jobs to be scheduled on $m$ identical machines, $f$ of which are prone to crashes by an adversary. A machine can inform other machines when a job is completed via the channel without collision detection. Performance is measured by the total number of available machine steps during the whole execution. Our goal is to study the impact of preemption (i.e., interrupting the execution of a job and resuming later in the same or differ…
With the rapidly growing demand of graph processing in the real scene, they have to efficiently handle massive concurrent jobs. Although existing work enable to efficiently handle single graph processing job, there are plenty of memory access redundancy caused by ignoring the characteristic of data access correlations. Motivated such an observation, we proposed two-level scheduling strategy in this paper, which enables to enhance the efficiency of data access and to accelerate the convergence speed of concurrent jobs. Firstly, correlations-aware job scheduling allows concurrent jobs to process…
The current Cloud infrastructure services (IaaS) market employs a resource-based selling model: customers rent nodes from the provider and pay per-node per-unit-time. This selling model places the burden upon customers to predict their job resource requirements and durations. Inaccurate prediction by customers can result in over-provisioning of resources, or under-provisioning and poor job performance. Thanks to improved resource virtualization and multi-tenant performance isolation, as well as common frameworks for batch jobs, such as MapReduce, Cloud providers can predict job completion time…
Training machine learning (ML) models with large datasets can incur significant resource contention on shared clusters. This training typically involves many iterations that continually improve the quality of the model. Yet in exploratory settings, better models can be obtained faster by directing resources to jobs with the most potential for improvement. We describe SLAQ, a cluster scheduling system for approximate ML training jobs that aims to maximize the overall job quality. When allocating cluster resources, SLAQ explores the quality-runtime trade-offs across multiple jobs to maximize sys…
We study a difficult problem of how to schedule complex workflows with precedence constraints under a limited budget in the cloud environment. We first formulate the scheduling problem as an integer programming problem, which can be optimized and used as the baseline of performance. We then consider the traditional approach of scheduling jobs in a prioritized order based on the upward-rank of each job. For those jobs with no precedence constraints among themselves, the plain upward-rank priority scheme assigns priorities in an arbitrary way. We propose a job prioritization scheme that uses Mar…
Increasing data volumes in scientific experiments necessitate the use of high-performance computing (HPC) resources for data analysis. In many scientific fields, the data generated from scientific instruments and supercomputer simulations must be analyzed rapidly. In fact, the requirement for quasi-instant feedback is growing. Scientists want to use results from one experiment to guide the selection of the next or even to improve the course of a single experiment. Current HPC systems are typically batch-scheduled under policies in which an arriving job is run immediately only if enough resourc…
Given a system model where machines have distinct speeds and power ratings but are otherwise compatible, we consider various problems of scheduling under resource constraints on the system which place the restriction that not all machines can be run at once. These can be power, energy, or makespan constraints on the system. Given such constraints, there are problems with divisible as well as non-divisible jobs. In the setting where there is a constraint on power, we show that the problem of minimizing makespan for a set of divisible jobs is NP-hard by reduction to the knapsack problem. We then…
Today's clusters often have to divide resources among a diverse set of jobs. These jobs are heterogeneous both in execution time and in their rate of arrival. Execution time heterogeneity has lead to the development of hybrid schedulers that can schedule both short and long jobs to ensure good task placement. However, arrival rate heterogeneity, or burstiness, remains a problem in existing schedulers. These hybrid schedulers manage resources on statically provisioned cluster, which can quickly be overwhelmed by bursts in the number of arriving jobs. In this paper we propose CloudCoaster, a hyb…
The problem of attaining energy efficiency in distributed systems is of importance, but a general, non-domain-specific theory of energy-minimal scheduling is far from developed. In this paper, we classify the problems of energy-minimal scheduling and present theoretical foundations of the same. We derive results concerning energy-minimal scheduling of independent jobs in a distributed system with functionally similar machines with different working and idle power ratings. The machines considered in our system can have identical as well as different speeds. If the jobs can be divided into arbit…
In this paper, we propose Peacock, a new distributed probe-based scheduler which handles heterogeneous workloads in data analytics frameworks with low latency. Peacock mitigates the \emph{Head-of-Line blocking} problem, i.e., shorter tasks are enqueued behind the longer tasks, better than the state-of-the-art. To this end, we introduce a novel probe rotation technique. Workers form a ring overlay network and rotate probes using elastic queues. It is augmented by a novel probe reordering algorithm executed in workers. We evaluate the performance of Peacock against two state-of-the-art probe-bas…
Motivated by applications in grid computing and project management, we study multiprocessor scheduling in scenarios where there is uncertainty in the successful execution of jobs when assigned to processors. We consider the problem of multiprocessor scheduling under uncertainty, in which we are given n unit-time jobs and m machines, a directed acyclic graph C giving the dependencies among the jobs, and for every job j and machine i, the probability p_{ij} of the successful completion of job j when scheduled on machine i in any given particular step. The goal of the problem is to find a schedul…
We consider a multi-organizational system in which each organization contributes processors to the global pool but also jobs to be processed on the common resources. The fairness of the scheduling algorithm is essential for the stability and even for the existence of such systems (as organizations may refuse to join an unfair system). We consider on-line, non-clairvoyant scheduling of sequential jobs. The started jobs cannot be stopped, canceled, preempted, or moved to other processors. We consider identical processors, but most of our results can be extended to related or unrelated processors…
Recent years have witnessed a large amount of decentralized data in multiple (edge) devices of end-users, while the aggregation of the decentralized data remains difficult for machine learning jobs due to laws or regulations. Federated Learning (FL) emerges as an effective approach to handling decentralized data without sharing the sensitive raw data, while collaboratively training global machine learning models. The servers in FL need to select (and schedule) devices during the training process. However, the scheduling of devices for multiple jobs with FL remains a critical and open problem. …
Modern data centers serve workloads which are capable of exploiting parallelism. When a job parallelizes across multiple servers it will complete more quickly, but jobs receive diminishing returns from being allocated additional servers. Because allocating multiple servers to a single job is inefficient, it is unclear how best to allocate a fixed number of servers between many parallelizable jobs. This paper provides the first optimal allocation policy for minimizing the mean slowdown of parallelizable jobs of known size when all jobs are present at time 0. Our policy provides a simple closed …
We propose constant approximation algorithms for generalizations of the Flexible Flow Shop (FFS) problem which form a realistic model for non-preemptive scheduling in MapReduce systems. Our results concern the minimization of the total weighted completion time of a set of MapReduce jobs on unrelated processors and improve substantially on the model proposed by Moseley et al. (SPAA 2011) in two directions. First, we consider each job consisting of multiple Map and Reduce tasks, as this is the key idea behind MapReduce computations, and we propose a constant approximation algorithm. Then, we int…
With the growing constraints on power budget and increasing hardware failure rates, the operation of future exascale systems faces several challenges. Towards this, resource awareness and adaptivity by enabling malleable jobs has been actively researched in the HPC community. Malleable jobs can change their computing resources at runtime and can significantly improve HPC system performance. However, due to the rigid nature of popular parallel programming paradigms such as MPI and lack of support for dynamic resource management in batch systems, malleable jobs have been largely unrealized. In t…
Nowadays large-scale distributed machine learning systems have been deployed to support various analytics and intelligence services in IT firms. To train a large dataset and derive the prediction/inference model, e.g., a deep neural network, multiple workers are run in parallel to train partitions of the input dataset, and update shared model parameters. In a shared cluster handling multiple training jobs, a fundamental issue is how to efficiently schedule jobs and set the number of concurrent workers to run for each job, such that server resources are maximally utilized and model training can…
Scheduling is an important task allowing parallel systems to perform efficiently and reliably. For modern computation systems, divisible load is a special type of data which can be divided into arbitrary sizes and independently processed in parallel. Such loads are commonly encountered in applications which are processing a great amount of similar data units. For a multi-task processor, the processor's speed may be time-varying due to the arrival and departure of other background jobs. This paper studies an optimal divisible loads scheduling problem on a single level tree network, whose proces…
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…
Applications in data-parallel computing typically consist of multiple stages. In each stage, a set of intermediate parallel data flows (Coflow) is produced and transferred between servers to enable starting of next stage. While there has been much research on scheduling isolated coflows, the dependency between coflows in multi-stage jobs has been largely ignored. In this paper, we consider scheduling coflows of multi-stage jobs represented by general DAGs (Directed Acyclic Graphs) in a shared data center network, so as to minimize the total weighted completion time of jobs. This problem is sig…
Realizing an optimal task scheduling by taking into account the business importance of jobs has become a matter of interest in pay and use model of Cloud computing. Introduction of an appropriate model for an efficient task scheduling technique could derive benefit to the service providers as well as clients. In this paper, we have addressed two major challenges which has implications on the performance of the Cloud system. One of the major issues is handling technical aspects of distributing the tasks for targeted gains and the second issue is related to the handling of the business priority …
Scheduling questions arise naturally in many different areas among which operating system design, compiling,... In real life systems, the characteristics of the jobs (such as release time and processing time) are usually unknown and unpredictable beforehand. The system is typically unaware of the remaining work in each job or of the ability of the job to take advantage of more resources. Following these observations, we adopt the job model by Edmonds et al (2000, 2003) in which the jobs go through a sequence of different phases. Each phase consists of a certain quantity of work and a speed-up …
