Intelligent Resource Scheduling at Scale: A Machine Learning Perspective

Abstract

Resource scheduling refers to the problem of packing tasks with multi-dimensional resource requirements and non-functional constraints. The exhibited heterogeneity of workload and server characteristics in Cloud-scale or Internetscale environments has raised unprecedented new challenges for cluster scheduling. Compared with ad-hoc heuristics for a multi-resource cluster scheduling problem, machine learning (ML) approaches can in turn facilitate improved efficiency in resource management. In this paper, we describe and discuss how ML can be used to autonomously exploit and understand both workloads and environments, and to learn how to efficiently deal with scheduling problems such as consolidating co-located workloads, handling resource requests, guaranteeing applications' QoS, and mitigating tailed stragglers etc. The scheduling procedure can be fundamentally learned from experience rather than interference by human subjectivity. Additionally, we present a generalized ML-based solution and demonstrate its effectiveness through a case study of the performance-centric node classification and straggler mitigation method. We believe that the rethinking from a ML perspective can steer the architecture optimization and efficiency improvement.

TOPOSCH: Latency-Aware Scheduling Based on Critical Path Analysis on Shared YARN Clusters

“…Multiple transactions within a DLRA have strong dependencies across multiple microservices. Load variability, however, indicates a temporal-spatio behaviors over time and across nodes [1] [2][3] [4]. A user request (e.g., an application request, a database query, a file access operation) will transverse a collection of microservices before being responded.…”

Energy saving scheduling in a fog-based IoT application by Bayesian taskclassification approach

“…For the scenarios in which the learning data will be produced slowly and the states are countable, we can use some type of supervised learning such as classification (discrete variables) or regression (continuous variables) [19]. According to [20] the impacts of using the ML concepts such as the mechanism of automatic resource allocation, scheduling, and the smart management of resources on cloud environments have been studied.…”

“…According to [19], in comparison with the ad hoc heuristic, the approaches of ML can be helpful through intelligent resource allocation, selection of action according to the conceptual states and environmental factors for scheduling. These approaches can represent a solution based on ML by modeling supervised learning and prepare the architecture.…”

“…Neural networks are compatible with the datacenter management and decrease the cost of cooling process of Google datacenters up to 40%. Deep learning (DL) at the edge of network decrease the amount of back propagated data to datacenters [19]. Efficiency of energy, spectrum, delay, stability, and safety are the key parameters which are considered at the stage of networking operation.…”

Investigation of the Possibility of Using Linear Regression for Predicting Memory Consumption in a Highload Information System

“…Ключевые слова: машинное обучение, линейная регрессия, процессы операционной системы, оперативная память. планирования и распределения ресурсов высоконагруженных систем (ВС) можно обойтись меньшими мощностями [3]. Ряд работ исследует возможность предсказания краткосрочной нагрузки для эффективной работы энергосистемы с помощью внешних параметров (температура воздуха, осадки, скорость ветра и прочее) [4][5][6].…”