Home // ICNS 2021, The Seventeenth International Conference on Networking and Services // View article

CurTail: Distributed Cotask Scheduling with Guaranteed Tail-Latency SLO

Authors:
Zhijun Wang
Hao Che
Hong Jiang

Keywords: Tail latency SLO guarantee, cotask scheduling, datacenter.

Abstract:
Today's user-facing interactive datacenter services, such as web searching and social networking, have to meet stringent tail latency Service Level Objectives (SLOs). Unfortunately, due to the scale-out nature of the workloads, how to enable both tail-latency-SLO guarantee for such services and high resource utilization remains a critical challenge. In this paper, we propose a distributed Cotask scheduler with guaranteed Tail-latency SLO (CurTail), aiming at providing both job tail-latency-SLO guarantee and high resource utilization. CurTail is a top-down, holistic approach. It decouples an upper job-level design from a lower task-level design. At the job level, a decomposition technique is proposed to translate a given job tail-latency SLO into task-level performance budgets for tasks in a cotask, i.e., the collection of tasks spawned by a job. At the task level, the task budgets are translated into both task compute and networking resource demands, hence allowing for distributed cotask scheduling. The preliminary testing results based on simulation indicate that CurTail can indeed provide job tail-latency SLO guarantee at high resource utilization.

Pages: 40 to 45

Copyright: Copyright (c) IARIA, 2021

Publication date: May 30, 2021

Published in: conference

ISSN: 2308-4006

ISBN: 978-1-61208-853-2

Location: Valencia, Spain

Dates: from May 30, 2021 to June 3, 2021