Graph partitioning algorithms for optimizing software deployment in mobile cloud computing

doi:10.1016/j.future.2012.07.003

Future Generation Computer Systems

Volume 29, Issue 2, February 2013, Pages 451-459

https://doi.org/10.1016/j.future.2012.07.003 Get rights and content

Abstract

As cloud computing is gaining popularity, an important question is how to optimally deploy software applications on the offered infrastructure in the cloud. Especially in the context of mobile computing where software components could be offloaded from the mobile device to the cloud, it is important to optimize the deployment, by minimizing the network usage. Therefore we have designed and evaluated graph partitioning algorithms that allocate software components to machines in the cloud while minimizing the required bandwidth. Contrary to the traditional graph partitioning problem our algorithms are not restricted to balanced partitions and take into account infrastructure heterogenity. To benchmark our algorithms we evaluated their performance and found they produce 10%–40% smaller graph cut sizes than METIS 4.0 for typical mobile computing scenarios.

Highlights

► Algorithms for partitioning software on the cloud are presented. ► KL-based algorithm allows fast partitioning for realtime use. ► Simulated annealing improves solution quality at the cost of computation capacity. ► Hybrid approach combines both. ► Comparison to METIS shows our algorithms find 10%–40% better graph cuts.

Introduction

Nowadays, the emergence of cloud computing is leading to a new paradigm of utility computing [1], where computing power is offered on an on-demand basis. Users are able to access applications, storage and processing over the Internet, via services offered by cloud providers on a pay-as-you-use scheme. The advantages for the end users are reduced cost, higher scalability and improved performance in comparison to maintaining their own private computer systems, dimensioned for peak load conditions. Moreover the elasticity of the cloud reduces the risks of overprovisioning (underutilization) or underprovisioning (saturation) [2].

The usage of the cloud is not only beneficial for web-based applications, but can also be used for other applications composed of many service components following the service-oriented programming paradigm. Some of these service components may have high needs regarding CPU power or memory consumption, and should therefore be executed on dedicated server machines in the cloud rather than on a regular desktop PC or mobile terminal, for example recognition components in an object recognition or speech to text application.

The adoption of the cloud paradigm poses the problem where to deploy software components, given the many options in terms of available hardware nodes in even moderate scale data centers. This deployment optimization is important to both the cloud user and the cloud provider in order to reduce costs. All components need to be deployed on a machine with sufficient CPU power while the communication overhead between different machines is preferably minimized as this introduces extra latency and network load. This problem can be modelled as a graph partitioning problem where a weighted graph of software components has to be partitioned in a number of parts representing the available machines. Moreover the optimal deployment can change over time, and thus in order to realise an optimal deployment a fast algorithm is desired.

An important scenario in this respect is cloud overspilling. In this scenario, a company offloads work from its own private infrastructure to a public cloud infrastructure on peak moments, as shown in Fig. 1. This enables the company to dimension its infrastructure for the average workload instead of the peak workload, reducing the underutilization and the cost of the private infrastructure. This situation is typical in digital document processing where one faces strict month-end or year-end deadlines, and thousands of batches of documents are to be processed. Typically, document processing systems support workflows consisting of a few tens of components (including content ingest, reformatting, layouting, merging, versioning, logging, output generation and printing components). As many of these components come in different versions, and potentially need to be instantiated for each customer separately, the number of components in such a scenario quickly amounts to a few hundreds.

A second use case is situated in the area of integrated simulation tools for engineering purposes [3]. These integrated tools typically involve multi-physics simulation (e.g. structural analysis, acoustic simulation and engine dynamics simulation in the case of designing a new automobile engine), and the number of simulation tools involved can easily amount to 10–20 for small engineering projects to a few 100 individually deployable components for a realistic engineering project. In such engineering endeavours, often parameter sweeps are executed to optimize the design (or to assess the sensitivity of the resulting design performance w.r.t. these parameters), necessitating multiple instances of these simulation components running concurrently, in order to arrive at realistic design times. Again, we end up with a component graph containing a few 100–1000 components.

This overspilling problem also arises in the context of mobile computing, where the cloud can be used to enhance the capabilities of a mobile device. Due to the restricted CPU power of mobile devices, the idea is to offload parts of the application at runtime to a cloud infrastructure [4]. The question then is which parts to offload and to deploy on which machines in the cloud–and how many–in order to spread the load while keeping the needed bandwidth low. In this case, the complexity of the partitioning problem depends on the granularity of the offloading, as offloading can be done on component [4], Java class [5] or method [6] level, resulting in graphs of tens, hundreds or thousands of components.

A special case of deployment optimization occurs when multiple mobile devices connected via a wireless network share their resources in order to enhance the user experience of all users, in a so called “cloudlet” [7], as shown in Fig. 2. This is the case when no internet uplink is available for offloading to the cloud, or when cloud offloading is not beneficial due to a high WAN latency. Because the bandwidth is a scarce resource and shared between all devices in the wireless LAN, a global optimization is needed taking into account all application components of all devices.

As an use case, we mention a mobile augmented reality application. When casting such an application into a component framework, the number of independently deployable components amounts to 5–10 [7], with different components for tracking camera movements, building a 3D map of the environment, recognizing objects, detecting collisions between objects, rendering a 3D overlay, etc. Other applications, such as 3D games, are reported to consist of 10–20 components [8]. Assuming a few tens of users connected to the same cloudlet and hence sharing computing and network resources, the number of components easily exceeds 100. As these users are connected through heterogeneous devices (a mix of low-end and high-end devices), an optimal deployment guaranteeing a minimal quality of experience for all users should be aimed for.

In this paper we present algorithms to partition a software application, composed of a number of components, on a number of interconnected machines in the cloud with different capacities while minimizing the communication cost between the components. In Section 2 related work regarding graph partitioning and task allocation on the grid is discussed. Section 3 more formally describes our problem and in Section 4 algorithms are proposed that solve the problem. In Section 5 the different algorithms are evaluated and compared regarding solution quality and execution time, and the influence of different parameters is discussed. In view of the use cases mentioned in this introduction, we focus on graphs containing 100–1000 components for this evaluation. We also compare our solutions to METIS 4.0 for partitioning graphs in $k$ balanced partitions and show the applicability of our algorithms in the mobile offloading scenario. Finally Section 6 concludes this paper.

Section snippets

Graph partitioning

Graph partitioning is a fundamental problem in many domains of computer science, such as VLSI design [9], parallel processing [10] and load balancing [11]. The graph partitioning problem tackles the problem of dividing a graph in $k$ equal sets while minimizing the edges between the sets. When $k = 2$ this is also referred to as the min-cut bipartitioning problem. Finding a good solution for this problem is known to be NP-Hard [12]. In the following we give a brief overview of the state-of-the-art

Problem statement

Let $G = (V, E)$ be an undirected graph where $V = {v_{1}, v_{2}, \dots, v_{N}}$ is a set of $N$ vertices and $E$ is a set of edges connecting the vertices. The vertices represent units of deployment in a distributed software system and the edges represent communication overhead between those units. Each vertex $v_{i}$ is assigned a cost $w_{i}$ that indicates the amount of resources (i.e. CPU power) this component needs. $C = (c_{i j})$ is the adjacency matrix of $G$ , i.e. if there exists an edge between $v_{i}$ and $v_{j}$ then $c_{i j}$ equals the weight

Integer linear programming (ILP)

The problem defined in Section 3 can be seen as an Integer Linear Programming (ILP) problem, thus an ILP solver (IBM ILOG CPLEX [37]) can be used to determine the optimal solution for this problem. The amount of time and resources needed to solve this problem grows exponentially with the size of the graph, so heuristics are needed to find a good solution faster. We can still use the ILP solution for smaller graphs to benchmark our algorithms that exhibit better scaling behaviour, possibly at

Evaluation results

To evaluate our algorithms we generated test graphs with different node sizes. Graphs were generated with the Eppstein power law generator [42] and the weights of the nodes and edges are assigned according an exponential distribution with parameter $λ$ equal to 0.1 and 0.005 respectively. Although the actual graph structure of an application will depend on the software design, we chose these graph configurations as most design principles aim to reduce the dependencies between different software

Conclusions and future work

In this paper we presented algorithms for partitioning software graphs for deployment on the cloud. The best partition is calculated taking into account the infrastructure heterogenity. In contrast to deployment optimization for computational grids, we do not minimize the execution time of a set of mesh structured tasks, but we focus on minimizing the bandwidth between software components. A multilevel KL-based algorithm is presented as a fast partitioner which allows realtime deployment

Tim Verbelen received his M.Sc. degree in Computer Science from Ghent University, Belgium in June 2009. Since then he has been working as a Ph.D. researcher at the Department of Information Technology (INTEC) of the Faculty of Engineering at Ghent University. His main research interests include mobile computing, context-aware computing and adaptive software.

References (42)

R. Buyya et al.
Cloud computing and emerging IT platforms: vision, hype, and reality for delivering computing as the 5th utility
Future Generation Computer Systems
(2009)
T. Verbelen et al.
Dynamic deployment and quality adaptation for mobile augmented reality applications
Journal of Systems and Software
(2011)
S. Ou et al.
An effective offloading middleware for pervasive services on mobile devices
Pervasive and Mobile Computing
(2007)
S. Han et al.
A resource aware software partitioning algorithm based on mobility constraints in pervasive grid environments
Future Generation Computer Systems
(2008)
C. Alpert
Recent directions in netlist partitioning: a survey
Integration, The VLSI Journal
(1995)
B. Hendrickson
Graph partitioning models for parallel computing
Parallel Computing
(2000)
H. Meyerhenke et al.
Graph partitioning and disturbed diffusion
Parallel Computing
(2009)
T.N. Bui et al.
Finding good approximate vertex and edge partitions is NP-hard
Information Processing Letters
(1992)
M.-W. Park et al.
A systematic procedure for setting parameters in simulated annealing algorithms
Computers & Operations Research
(1998)
M. Armbrust, A. Fox, R. Griffith, A.D. Joseph, R.H. Katz, A. Konwinski, G. Lee, D.A. Patterson, A. Rabkin, M. Zaharia,...

Noesis solutions, optimus....

M. Kristensen, Scavenger: transparent development of efficient cyber foraging applications, in: 2010 IEEE International...

T. Verbelen, P. Simoens, F. De Turck, B. Dhoedt, Cloudlets: bringing the cloud to the mobile user, in: Proc. of the 3rd...

P.-O. Fjallstrom

Algorithms for graph partitioning: a survey

Computer and Information Science

(1998)

K. Schloegel et al.

The Sourcebook of Parallel Computing

(2003)

B. Kernighan et al.

An efficient heuristic procedure for partitioning graphs

Bell System Technical Journal

(1970)

C. Fiduccia et al.

A linear-time heuristic for improving network partitions

L. Sun et al.

An effective multi-level algorithm based on simulated annealing for bisecting graph

Lecture Notes in Computer Science

(2007)

L. Sun et al.

A new multi-level algorithm based on particle swarm optimization for bisecting graph

Lecture Notes in Computer Science

(2007)

M. Leng et al.

An effective multi-level algorithm based on ant colony optimization for bisecting graph

Lecture Notes in Computer Science

(2007)

B. Hendrickson, R. Leland, A multilevel algorithm for partitioning graphs, in: Proceedings of the 1995 ACM/IEEE...

Cited by (72)

Makespan-minimization workflow scheduling for complex networks with social groups in edge computing
2020, Journal of Systems Architecture
Edge computing enables users to offload certain computation loads in complex applications to a nearby network consisting of multiple mobile devices such that the mobile devices’ resources can be integrated to afford these complex applications. Due to the social relationships among the owners of mobile devices, networks in edge computing are always complex and consist of multiple sub-networks intersecting at several joint devices. In such complex networks, a joint device can communicate with all devices belonging to all sub-networks that intersect at this joint device while a general device can only communicate with the devices belonging to the same sub-network. This paper studies the makespan-minimization workflow scheduling problem for the aforementioned complex networks in edge computing environments, and formulates the problem as an integer program. Due to the limited energy capacity of each device, the dependence among workflow tasks, as well as network complexity, it is challenging to achieve feasible scheduling solutions for the concerned problem. Therefore, we propose a family of task allocation strategies to cope with different types of tasks in the workflow. These strategies are further integrated with a greedy strategy to construct an improved greedy search (IGS) algorithm which is capable of generating feasible solutions satisfying all constraints. In addition, we propose an improved composite heuristic (ICH) algorithm that employs IGS to initialize a feasible solution and uses a two-layer improvement scheme to further enhance the quality of the initial solution. Simulation results show that our proposed IGS achieves an 100% probability of generating feasible solutions, as compared to a 2.3% probability achieved by using the general round-robin scheduling algorithm. Furthermore, IGS and ICH outperform the counterpart scheduling algorithms in terms of makespan reduction.
Secure software placement and configuration
2020, Future Generation Computer Systems
Citation Excerpt :
Mixed Integer (quadratically constrained) Quadratic Programming involves formulating the problem using a finite set of – real-valued or integer-valued – variables, and a finite set of constraints, which are equations or inequalities comprised of linear or quadratic expressions of the variables. It is more common to use Mixed Integer Linear Programming, in which the constraints are all linear, to solve application placement problems [22,23], but as we will see, we need the higher expressive power of quadratic constraints to express the more complex relationships in connection with the security controls. The steps of the proposed algorithm are summarized in Algorithm 1, and explained in the following.
Finding the best way to place the components of an application on a set of heterogeneous servers is a challenging task, especially if some components are associated with security requirements. To address security requirements, several security controls may be available, some of them software-based (e.g., encryption), others hardware-based (e.g., trusted execution environments). Security controls may incur widely varying performance overhead. There is a non-trivial interplay between application placement (which component to place on which server) and the configuration of security controls (which security control to activate for which component). On the one hand, placing a component on a secure server may make it unnecessary to use software-based security controls for the component. On the other hand, the overhead of using a specific security control may increase the resource requirements of a component so that it does not fit onto its designated server. Therefore, this paper addresses the joint problem of application placement and configuration of security controls. We formalize the problem and use mixed integer quadratic programming to solve it. A case study is used to demonstrate that the proposed approach can automatically determine the placement and configuration of complex applications.
A smartphone perspective on computation offloading—A survey
2020, Computer Communications
Citation Excerpt :
Then, profiling is carried out on every component (module/bundle/method/service) composing those applications to measure necessary data. One of the most typical methods to analyze and model an application running on smartphones is to use a call graph, which is a Directed Acyclic Graph (DAG) that represents the relationship between the computational components of that program [150–159]. Normally, the nodes of the graph represent the procedures/functions of a program and the edges represent the communication/invocations between them.
Computation offloading has emerged as one of the promising approaches to address the issues of restricted resources, leading to poor user experiences on smartphones in terms of battery life and performance when executing CPU intensive tasks. Meanwhile, an entire research domain has been initiated around this topic and several concepts have been studied touching both the smartphone and the cloud side. In this paper, we develop a categorization of fundamental aspects regarding computation offloading in heterogeneous cloud computing from the perspective of smartphone applications. We refer to heterogeneity in terms of the multitude of smartphone applications, the various uplink channels, and the variety of cloud solutions. We also survey state-of-the-art solutions for the identified categories. Finally, we conclude with a summary of the most important research challenges in making computation offloading reality.
Isolating critical flow path and algorithmic partitioning of the AND/OR mobile workflow graph
2020, Future Generation Computer Systems
The standard workflow modeling approaches have insufficient information about structural stability to visualize the tasks in a complex workflow graph. The hybrid AND/OR mobile workflow graph (MWG) model addresses this issue by visualizing a fully conditioned workflow graph for dynamic networks. However, the scheduling and mapping of large scale workflows to heterogeneous resources is NP-complete in parallel and distributed computing systems. This paper aims to partition large scale workflows into mappable sub-workflows by isolating the critical flow paths using a filter matrix method. A set of algorithms are designed to perform decomposability checking and partitions of a AND/OR mobile workflow graph. The filter matrix allows input of user’s constraints prior to the partitioning of a complex workflow graph. The experimental results show that valid partitions are obtained from various AND/OR mobile workflow graph. Moreover, non-linear regression analysis is presented to project the algorithmic time complexity for the partitioning of the large scale workflow graphs. A detailed comparative analysis of the proposed algorithm is included in the paper.
Service deployment and scheduling for improving performance of composite cloud services
2019, Computers and Electrical Engineering
Cloud computing and services have greatly changed the way how people develop and use software, and also raised a lot of new research issues. In this paper, we investigate two such important issues, service deployment and service request scheduling, for composite cloud services in dynamic cloud environments. We present load-aware service deployment approaches for dynamic workload and a service request scheduling method based on task ranking mechanisms. The proposed approaches can effectively improve the execution performance of composite cloud services in dynamic cloud environments. A series of simulation experiments have been conducted to evaluate the proposed approaches. The experimental results show that our approaches outperform previous methods significantly in terms of average service response time.
Internet of Things (IoT), mobile cloud, cloudlet, mobile IoT, IoT cloud, fog, mobile edge, and edge emerging computing paradigms: Disambiguation and research directions
2019, Journal of Network and Computer Applications
Currently, we are experiencing a technological shift, which is expected to change the way we program and interact with the world. Cloud computing and mobile computing are two prominent research areas that have already had such an impact. The Internet of Things (IoT), which is concerned with building a network of Internet-enabled devices to promote a smart environment, is another promising area of research. Numerous emerging computing paradigms related to those areas of research and/or their intersections have come into play. These include Mobile Cloud Computing (MCC), cloudlet computing, mobile clouds, mobile IoT computing, IoT cloud computing, fog computing, Mobile Edge Computing (MEC), edge computing, the Web of Things (WoT), the Semantic WoT (SWoT), the Wisdom WoT (W2T), opportunistic sensing, participatory sensing, mobile crowdsensing, and mobile crowdsourcing. Unfortunately, those paradigms suffer from the lack of standard definitions, and so we frequently encounter a single term referring to various paradigms or several terms referring to a single paradigm. Accordingly, this paper attempts to disambiguate those paradigms and explain how and where they fit in the above three areas of research and/or their intersections before it becomes a serious problem. They are tracked back to their inception as much as possible. This is in addition to discussing research directions in each area. The paper also introduces technologies related to the IoT such as ubiquitous and pervasive computing, the Internet of Nano Things (IoNT), and the Internet of Underwater Things (IoUT).

View all citing articles on Scopus

Tim Stevens received his M.Sc. degree in Computer Science from Ghent University, Belgium, in June 2001. In January 2009, he obtained his Ph.D. degree in Computer Science Engineering from the same university. Before joining the Department of Information Technology (INTEC) of Ghent University in August 2003, Tim was a database and system administrator for the Flemish public broadcasting company (VRT). At present, he is a postdoctoral researcher affiliated with INTEC.

His main research interests include future access network architectures, IPv6, Quality of Service (QoS) and traffic engineering in IP networks, anycast-based services, and cloud and grid computing.

Filip De Turck received his M.Sc. degree in Electronic Engineering from the Ghent University, Belgium, in June 1997. In May 2002, he obtained the Ph.D. degree in Electronic Engineering from the same university. During his Ph.D. research he was funded by the F.W.O.-V., the Fund for Scientific Research Flanders. From October 2002 until September 2008, he was a post-doctoral fellow of the F.W.O.-V. and part time professor, affiliated with the Department of Information Technology of the Ghent University. At the moment, he is a full-time professor affiliated with the Department of Information Technology of the Ghent University and the IBBT (Interdisciplinary Institute of Broadband Technology Flanders) in the area of telecommunication and software engineering. Filip De Turck is author or co-author of approximately 250 papers published in international journals or in the proceedings of international conferences. His main research interests include scalable software architectures for telecommunication network and service management, cloud computing, performance evaluation and design of new telecommunication and eHealth services.

Bart Dhoedt received a Masters degree in Electro-technical Engineering (1990) from Ghent University. His research, addressing the use of micro-optics to realize parallel free space optical interconnects, resulted in a Ph.D. degree in 1995. After a 2-year post-doc in opto-electronics, he became Professor at the Department of Information Technology.

Bart Dhoedt is responsible for various courses on algorithms, advanced programming, software development and distributed systems. His research interests include software engineering, distributed systems, mobile and ubiquitous computing, smart clients, middleware, cloud computing and autonomic systems. He is author or co-author of more than 300 publications in international journals or conference proceedings.

View full text

Graph partitioning algorithms for optimizing software deployment in mobile cloud computing

Abstract

Highlights

Introduction

Section snippets

Graph partitioning

Problem statement

Integer linear programming (ILP)

Evaluation results

Conclusions and future work

Future Generation Computer Systems

Journal of Systems and Software

Pervasive and Mobile Computing

Future Generation Computer Systems

Integration, The VLSI Journal

Parallel Computing

Parallel Computing

Information Processing Letters

Computers & Operations Research

Algorithms for graph partitioning: a survey

Computer and Information Science

The Sourcebook of Parallel Computing

An efficient heuristic procedure for partitioning graphs

Bell System Technical Journal

A linear-time heuristic for improving network partitions

An effective multi-level algorithm based on simulated annealing for bisecting graph

Lecture Notes in Computer Science

A new multi-level algorithm based on particle swarm optimization for bisecting graph

Lecture Notes in Computer Science

An effective multi-level algorithm based on ant colony optimization for bisecting graph

Lecture Notes in Computer Science