Elsevier

Performance Evaluation

Volume 91, September 2015, Pages 205-228
Performance Evaluation

Dynamic service migration and workload scheduling in edge-clouds

https://doi.org/10.1016/j.peva.2015.06.013Get rights and content

Abstract

Edge-clouds provide a promising new approach to significantly reduce network operational costs by moving computation closer to the edge. A key challenge in such systems is to decide where and when services should be migrated in response to user mobility and demand variation. The objective is to optimize operational costs while providing rigorous performance guarantees. In this paper, we model this as a sequential decision making Markov Decision Problem (MDP). However, departing from traditional solution methods (such as dynamic programming) that require extensive statistical knowledge and are computationally prohibitive, we develop a novel alternate methodology. First, we establish an interesting decoupling property of the MDP that reduces it to two independent MDPs on disjoint state spaces. Then, using the technique of Lyapunov optimization over renewals, we design an online control algorithm for the decoupled problem that is provably cost-optimal. This algorithm does not require any statistical knowledge of the system parameters and can be implemented efficiently. We validate the performance of our algorithm using extensive trace-driven simulations. Our overall approach is general and can be applied to other MDPs that possess a similar decoupling property.

Introduction

The increasing popularity of mobile applications (such as social networking and photo sharing) running on handheld devices is putting a significant burden on the capacity of cellular and backhaul networks. These applications are generally comprised of a front-end component running on the handheld and a back-end component (that performs data processing and computation) that typically runs on the cloud. While this architecture enables applications to take advantage of the on-demand feature of cloud computing, it also introduces new challenges in the form of increased network overhead and latency. A promising approach to address these challenges is to move such computation closer to the network edge. Here, it is envisioned that entities (such as basestations in a cellular network) closer to the network edge would host smaller-sized cloud-like infrastructure distributed across the network. This idea has been variously termed as Cloudlets  [1], Fog Computing  [2], Edge Computing  [3], and Follow Me Cloud  [4], to name a few. The trend towards edge-clouds is expected to accelerate as more users perform a majority of their computations on handhelds and as newer mobile applications get adopted.

One of the key design issues in edge-clouds is service migration: should a service currently running in one of the edge-clouds be migrated as the user locations change, and if yes, where? This question stems from the basic tradeoff between the cost of service migration vs. the reduction in network overhead and latency for users that can be achieved after migration. While conceptually simple, it is challenging to make this decision in an optimal manner because of the uncertainty in user mobility and request patterns. Because edge-clouds are distributed at the edge of the network, their performance is closely related to user dynamics. These decisions get even more complicated when the number of users and applications is large and there is heterogeneity across edge-clouds. Note that the service migration decisions affect workload scheduling as well (and vice versa), so that in principle these decisions must be made jointly.

The overall problem of dynamic service migration and workload scheduling to optimize system cost while providing end-user performance guarantees can be formulated as a sequential decision making problem in the framework of MDPs  [5], [6]. This approach, although very general, suffers from several drawbacks. First, it requires extensive knowledge of the statistics of the user mobility and request arrival processes that can be impractical to obtain in a dynamic network. Second, even when this is known, the resulting problem can be computationally challenging to solve. Finally, any change in the statistics would make the previous solution suboptimal and require recomputing the optimal solution.

In this paper, we present a new methodology that overcomes these drawbacks. Our approach is inspired by the technique of Lyapunov optimization  [7], [8] which is a general framework for designing optimal control algorithms for non-MDP problems without requiring any knowledge of the transition probabilities. Specifically, these are problems where the cost functions and control decisions are functionals of states that evolve independently of the control actions. However, as we will show later, this condition does not hold for the joint service migration and workload scheduling problem studied in this paper. A key contribution of this work is to develop a methodology that enables us to still apply the Lyapunov optimization technique to this MDP while preserving its attractive features.

Section snippets

Related work

The general problem of resource allocation and workload scheduling in cloud computing systems using the framework of stochastic optimization has been considered in several recent works. Specifically,  [9] considers a stochastic model for a cloud computing cluster, where requests for virtual machines (VMs) arrive according to a stochastic process. Each VM request is specified in terms of a vector of resources (such as CPU, memory and storage space) and its duration and must be placed on physical

Problem formulation

We consider an edge-cloud system comprised of M distributed edge-clouds and one back-end cloud that together host K applications (see Fig. 1). The system also consists of N users that generate application requests over time. The collection of edge and back-end clouds supports these applications by providing the computational resources needed to serve user requests. The users are assumed to be mobile while the edge and back-end clouds are static. We assume a time-slotted model and use the notion

MDP relaxation and decoupling

Consider a relaxation of the original MDP discussed in Section  3.2 where we replace the average delay constraints by the following queue stability constraints k,m. μ¯km+υ¯kmR¯kmϵifR¯km>0 where R¯km,μ¯km and υ¯km respectively denote the time-average expected arrival rate, local service rate and back-end routing rate under any control algorithm. It can be shown that meeting these constraints ensures that all queues are rate stable  [8]. Further, we add the constraints that C¯=c,E¯=e, and W¯=

Online control algorithm

We now present an online control algorithm that makes joint request routing and application configuration decisions as a function of the system state (l(t),a(t),h(t),U(t)). However, unlike traditional MDP solution approaches such as dynamic programming  [5], [6], this algorithm does not require any knowledge of the transition probabilities that govern the system dynamics. In addition to the request queues Ukm(t), for each (k,m) this algorithm maintains the following “delay-aware” queues that

Performance analysis

We now analyze the performance of the online control algorithm presented in Section  5. This is based on the technique of Lyapunov optimization over renewal periods  [8], [17] where we compare the ratio of a weighted combination of the Lyapunov drift and costs over a renewal period and the length of the period under the online algorithm with the same ratio under a stationary algorithm that is queue backlog independent. This stationary algorithm is defined similarly to the decoupled control

Evaluations

We evaluate the performance of our control algorithm using simulations. To show both the theoretical and real-world behaviors of the algorithm, we consider two types of user mobility traces. The first is a set of synthetic traces obtained from a random-walk user mobility model while the second is a set of real-world traces of San Francisco taxis  [21]. We assume that the edge-clouds are co-located with a subset of the basestations of a cellular network. A hexagonal symmetric cellular structure

Conclusions

In this paper, we have developed a new approach for solving a class of constrained MDPs that possess a decoupling property. When this property holds, our approach enables the design of simple online control algorithms that do not require any knowledge of the underlying statistics of the MDPs, yet are provably optimal. The resulting solution is markedly different from classical dynamic programming based approaches and does not suffer from the associated “curse of dimensionality” or convergence

Acknowledgments

This research was sponsored in part by the US Army Research Laboratory and the UK Ministry of Defence and was accomplished under Agreement Number W911NF-06-3-0001. The views and conclusions contained in this document are those of the author(s) and should not be interpreted as representing the official policies, either expressed or implied, of the US Army Research Laboratory, the US Government, the UK Ministry of Defence or the UK Government. The US and UK Governments are authorized to reproduce

Rahul Urgaonkar is a Research Staff Member with the Cloud-Based Networks group at the IBM T.J. Watson Research Center. He is currently a task lead on the US Army Research Laboratory (ARL) sponsored Network Science Collaborative Technology Alliance (NS CTA) program. He is also a Primary Researcher in the US/UK International Technology Alliance (ITA) research programs. His research is in the areas of stochastic optimization, algorithm design, and optimal control with applications to communication

References (22)

  • M. Satyanarayanan

    Cloudlets: At the leading edge of cloud-mobile convergence

  • F. Bonomi et al.

    Fog computing and its role in the Internet of things

  • S. Davy et al.

    Challenges to support edge-as-a-service

    IEEE Commun. Mag.

    (2014)
  • T. Taleb et al.

    Follow me cloud: interworking federated clouds and distributed mobile networks

    IEEE Netw.

    (2013)
  • D.P. Bertsekas

    Dynamic Programming and Optimal Control

    (2000)
  • M.L. Puterman

    Markov Decision Processes: Discrete Stochastic Dynamic Programming

    (1994)
  • L. Georgiadis et al.

    Resource allocation and cross-layer control in wireless networks

    Found. Trends Netw.

    (2006)
  • M.J. Neely

    Stochastic Network Optimization with Application to Communication and Queueing Systems

    (2010)
  • S. Maguluri, R. Srikant, L. Ying, Stochastic models of load balancing and scheduling in cloud computing clusters, in:...
  • Y. Guo, A. Stolyar, A. Walid, Shadow-routing based dynamic algorithms for virtual machine placement in a network cloud,...
  • Y. Guo et al.

    Online algorithms for joint application-VM-physical-machine auto-scaling in a cloud

  • Cited by (236)

    View all citing articles on Scopus

    Rahul Urgaonkar is a Research Staff Member with the Cloud-Based Networks group at the IBM T.J. Watson Research Center. He is currently a task lead on the US Army Research Laboratory (ARL) sponsored Network Science Collaborative Technology Alliance (NS CTA) program. He is also a Primary Researcher in the US/UK International Technology Alliance (ITA) research programs. His research is in the areas of stochastic optimization, algorithm design, and optimal control with applications to communication networks and cloud-computing systems. Before joining IBM research, Dr. Urgaonkar was a Scientist with the Network Research group at Raytheon BBN Technologies where he worked on several government funded projects, including the NS CTA and ITA programs. He obtained his Masters and Ph.D. degrees from the University of Southern California in 2005 and 2011 respectively and his Bachelor’s degree (all in Electrical Engineering) from the Indian Institute of Technology Bombay in 2002.

    Shiqiang Wang received the B.Eng. and M.Eng. degrees from Northeastern University, China, respectively in 2009 and 2011. He is currently working toward the Ph.D. degree at the Department of Electrical and Electronic Engineering, Imperial College London, United Kingdom. In the summers of 2014 and 2013, he was at IBM T.J. Watson Research Center, Yorktown Heights, NY, United States. In the autumn of 2012, he was at NEC Laboratories Europe, Heidelberg, Germany. His research interests include dynamic control mechanisms, optimization algorithms, protocol design and prototyping, with applications to mobile cloud computing, hybrid and heterogeneous networks, ad-hoc networks, and cooperative communications. He has over 20 scholarly publications, and has served as a technical program committee (TPC) member or reviewer for a number of international journals and conferences.

    Ting He received the B.S. degree in computer science from Peking University, China, in 2003 and the Ph.D. degree in electrical and computer engineering from Cornell University, Ithaca, NY, in 2007. In 2007, Ting joined the IBM T.J. Watson Research Center, where she is currently a Research Staff Member in the Network Analytics Research Group. At IBM, she has worked as a primary researcher and task lead in several research programs including the International Technology Alliance (ITA) program funded by US ARL and UK MoD, the ARRA program funded by NIST, and the Social Media in Strategic Communication (SMISC) program funded by DARPA. Her work is in the broad areas of network modeling, statistical inference, and information theory. He is a senior member of IEEE. She has served as the Membership co-chair of ACM N2Women and the TPC of a range of communications and networking conferences, including IEEE INFOCOM, IEEE SECON, IEEE/ACM IWQoS, IEEE MILCOM, IEEE ICNC, IFIP Networking, etc. She received the Outstanding Contributor Award from IBM Research in 2009 and 2013. She received the Best Paper Award at the 2013 International Conference on Distributed Computing Systems (ICDCS), a Best Paper Nomination at the 2013 Internet Measurement Conference (IMC), and the Best Student Paper Award at the 2005 International Conference on Acoustic, Speech and Signal Processing (ICASSP). In school, she was an Outstanding College Graduate of Beijing Area and an Outstanding Graduate of Peking University in 2003, and a winner of the Excellent Student Award of Peking University from 1999 to 2002.

    Murtaza Zafer received the B.Tech. degree in Electrical Engineering from the Indian Institute of Technology, Madras, in 2001, and the Ph.D. and S.M. degrees in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology in 2003 and 2007 respectively. He currently works at Nyansa Inc., where he leads the analytics portfolio of the company. Prior to this, he was a Senior Research Engineer at Samsung Research America, where his research focused on machine learning, deep-neural networks, big data and cloud computing. From 2007–2013 he was a Research Scientist at the IBM T.J. Watson Research Center, New York, where his research focused on computer and communication networks, data-analytics and cloud computing. He was a technical lead on several research projects in the US–UK funded multi-institutional International Technology Alliance program with emphasis on fundamental research in mobile wireless networks. He has previously worked at the Corporate R&D center of Qualcomm Inc. and at Bell Laboratories, Alcatel-Lucent Inc., during the summers of 2003 and 2004 respectively. Zafer serves as an Associate Editor for the IEEE Network magazine. He is a co-recipient of the Best Student Paper award at the International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt) in 2005, a recipient of the Siemens and Philips Award in 2001 and a recipient of several invention achievement awards at IBM Research.

    Kevin Chan is research scientist with the Computational and Information Sciences Directorate at the US Army Research Laboratory (Adelphi, MD). Previously, he was an ORAU postdoctoral research fellow at ARL (2008–10). His research interests are in network science, with past work in dynamic networks, trust and distributed decision making and quality of information through ARL’s Network Science Collaborative Technology Alliance and Network and Information Sciences International Technology Alliance. Prior to ARL, he received a Ph.D. in Electrical and Computer Engineering (ECE) and MSECE from Georgia Institute of Technology (Atlanta, GA) in 2003 and 2008, respectively. He also received a BS in ECE/EPP from Carnegie Mellon University (Pittsburgh, PA) in 2001.

    Kin K. Leung received his B.S. degree from the Chinese University of Hong Kong in 1980, and his M.S. and Ph.D. degrees from University of California, Los Angeles, in 1982 and 1985, respectively. He joined AT&T Bell Labs in New Jersey in 1986 and worked at its successors, AT&T Labs and Lucent Technologies Bell Labs, until 2004. Since then, he has been the Tanaka Chair Professor in the Electrical and Electronic Engineering (EEE), and Computing Departments at Imperial College in London. He is the Head of Communications and Signal Processing Group in the EEE Department. His current research focuses on protocols, optimization and modeling of various wireless networks. He also works on multi-antenna and cross-layer designs for these networks. He received the Distinguished Member of Technical Staff Award from AT&T Bell Labs (1994), and was a co-recipient of the Lanchester Prize Honorable Mention Award (1997). He was elected an IEEE Fellow (2001), received the Royal Society Wolfson Research Merits Award (2004–09) and became a member of Academia Europaea (2012). He also received several best paper awards, including the IEEE PIMRC 2012 and ICDCS 2013. He has actively served on conference committees. He serves as a member (2009–11) and the chairman (2012–15) of the IEEE Fellow Evaluation Committee for Communications Society. He was a guest editor for the IEEE JSAC, IEEE Wireless Communications and the MONET journal, and as an editor for the JSAC: Wireless Series, IEEE Transactions on Wireless Communications and IEEE Transactions on Communications. Currently, he is an editor for the ACM Computing Survey and International Journal on Sensor Networks.

    View full text