WIXLIB

2017 IEEE International Parallel and Distributed Processing SymposiumContainer-Based Cloud Platform for MobileComputation Offloading Song Wu , Chao Niu , Jia Rao† , Hai Jin and Xiaohai Dai Services Computing Technology and System Lab, Cluster and Grid Computing LabSchool of Computer Science and Technology, Huazhong University of Science and Technology Wuhan, 430074, ChinaEmail: {wusong, chaoniu, hjin, seafooler}@hust.edu.cn† The University of Texas at Arlington, USAEmail: jia.rao@uta.eduAbstract—With the explosive growth of smartphones andcloud computing, mobile cloud, which leverages cloud resourceto boost the performance of mobile applications, becomes attractive. Many efforts have been made to improve the performanceand reduce energy consumption of mobile devices by offloadingcomputational codes to the cloud. However, the offloading costcaused by the cloud platform has been ignored for many years.In this paper, we propose Rattrap, a lightweight cloud platformwhich improves the offloading performance from cloud side. Toachieve such goals, we analyze the characteristics of typical offloading workloads and design our platform solution accordingly.Rattrap develops a new runtime environment, Cloud AndroidContainer, for mobile computation offloading, replacing heavyweight virtual machines (VMs). Our design exploits the ideaof running operating systems with differential kernel featuresinside containers with driver extensions, which partially breaksthe limitation of OS-level virtualization. With proposed resourcesharing and code cache mechanism, Rattrap fundamentallyimproves the offloading performance. Our evaluation showsthat Rattrap not only reduces the startup time of runtimeenvironments and shows an average speedup of 16x, but alsosaves a large amount of system resources such as 75% memoryfootprint and at least 79% disk capacity. Moreover, Rattrapimproves offloading response by as high as 63% over the cloudplatform based on VM, and thus saving the battery life.cloud are the long VM startup time and high virtualizationoverhead. Interactivity and mobility are the keys to thesuccess of mobile cloud computing, but heavyweight VMsolutions cannot meet these requirements [5, 6]. Though prestarting VMs can reduce the VM startup time, it wouldinevitably incur high resource cost because the number ofoffloading requests is probably large [4].OS-level virtualization, a.k.a, container-based virtualization[7–10] has recently attracted much attention due to its nearnative performance and low virtualization overhead. Building a mobile cloud using container-based virtualization isa promising idea, but presents several challenges. First, thecontainer must run the same OS as the host. Much effortis needed to customize a container to support mobile apps.Second, OS containers offer less isolation between mobileOSes, making it possible to exploit optimizations betweenmobile apps to further improve performance. However, therelacks a comprehensive study of offloading performance toguide the optimizations.In this paper, we address the above two challenges anddesign Rattrap, a mobile offloading cloud platform based onOS containers. To support the execution of mobile apps inthe cloud, we develop Cloud Android Container, a mobileOS environment built directly on general purpose server OS.We perform a comprehensive study of offloading performance in representative mobile workloads, which motivatesthe development of two optimizations for mobile offloading:shared resource layer between multiple apps and mobilecode cache. Experimental results using LXC container andAndroid apps show significant performance improvement oncode offloading compared to VM-based cloud platforms. Thestartup time of the mobile OS is reduced from more than 28susing VMs to less than 2s with Cloud Android Container.Our container-based cloud saves as much as 75% memoryfootprint and at lease 79% disk usage by hosting the mobileOS in the cloud. Moreover, the performance of code executionhas been improved by as much as 40%. In summary, thispaper makes the following contributions: We discuss the behaviors of mobile cloud applications,characterize representative offloading workloads, andsummarize the problems current cloud platform faces.These provide the guidelines of how to make the cloudside better serve offloading requests.I. I NTRODUCTIONMobile cloud computing, which offloads computation onmobile devices to a cloud platform, can enable executionof computationally intensive applications on mobile deviceswith enhanced user experience.The challenges lie in how toseamlessly integrate mobile runtime environments into cloudplatforms and how to perform mobile computation offloadingin a cost-effective manner. Existing work mainly focused onthe design of code offloading frameworks and addressed theissues of code partitioning, offloading decision and methodology. The design of a cloud platform to support mobilecomputation offloading has largely been ignored. Previousstudies assumed that computation from the cloud is ubiquitousand always ready to use.Many practical issues arise when building a real mobilecloud. The cloud should provide on-demand execution environments for mobile computation codes, e.g., the Androidmobile OS. Many offloading studies [1–4] used virtual machines (VMs) to host mobile OS in the cloud. Hardwarevirtualization allows different guest OSes to co-exist in thecloud platform. The main drawbacks of using VMs in mobile1530-2075/17 31.00 2017 IEEEDOI 10.1109/IPDPS.2017.47123

We devise the idea of dynamically extending the hostOS kernel with mobile OS drivers and make the mobilecloud platform based on OS-level virtualization cometrue. By implementing drivers for various kernel features,containers can run certain different OSes, which partiallybreaks the kernel limitation of OS-level virtualization.With containers running mobile environments, we propose a set of techniques to improve its efficiency, including customization of mobile OS, sharing commonresources, and code cache mechanisms, which makes ourwork no longer a simple code runtime environment.III. W ORKLOAD A NALYSISIn this section, we perform a characterization of benchmarkapplications in used in previous mobile offloading research,and discuss the problems of existing cloud platforms based onVM. As will be discussed later in Section IV, these findingsabout offloading workloads motivated our system design.A. Experiment SetupIn this series of experiments, four representative Androidapplications of four different categories, which have beenwidely adopted by previous researches [1, 12–14], are usedas our benchmark workloads. Image tools are the most common benchmarks usedin previous researches and represent the computationintensive workloads with file transfer. OCR (OpticalCharacter Recognition) is based on the Google Tesseractlibrary whose real computation is implemented with JavaNative Interface (JNI) code written in C . Games interact with user continually, representing workloads with intensive network communications. ChessGame is an Android port of the CuckooChess Enginewhich ranks top 200 in Computer Chess Rating Lists40/401 . Anti-Virus has been adopted increasingly since the rapidmalware emergence. VirusScan checks the target withvirus database search and spawns more I/O requests thanother benchmarks. Mathematical tools, like Linpack, are implemented inordinary Android Java. They are often used to representpure computation.The client applications run on 5 Android devices and usethe Java reflection techniques to enable the offloading ofcomputation codes. The mobile code runtime environmentson the cloud is based on Android-x86 [16] VM. More detailsabout experimental settings, like the configuration of cloudservers, can be found in Section VI.II. BACKGROUNDA. Basic Offloading MechanismWith the rise of mobile cloud computing, the idea ofoffloading computation has been a research hotspot for a longtime. Unlike existing mobile applications which run locallyand directly request data from content providers, mobiledevices can offload parts of the workloads to the cloud,making cloud a powerful worker to process computationtasks. Generally, a basic offloading system [2, 3, 11–14]is composed of two parts. The client side runs on mobiledevices, which controls offloading computational code to thecloud when needed. The cloud side is responsible for handlingoffloading requests and provides runtime environments.Much research effort has been duplicating mobile runtimeenvironments in the cloud to support offloading. To achievethis, many frameworks [2, 3] leverage the virtualizationtechnology, namely mobile OS virtual machine. The cloudplatforms in these frameworks have nearly the same runtimeenvironments as mobile devices, which makes it straightforward to run mobile code on the cloud side. However, to usea VM instance, the cloud has to install and boot a guest OSin the VM, which incurs substantial delay. For example, aStandard Medium instance (m1.medium) on Amazon EC2 hasan average setup time of 28.4 seconds, which is undesirablefor mobile requests. Therefore, on-demand deploying VMs isan expensive operation and not suitable for real-time mobileoffloading [6].B. Performance Penalty of Runtime PreparationAs described in [6], the key reason that existing cloudapproaches are not suitable for mobile cloud applications isthe unacceptable long startup time of VMs. In this section, weadopt all 4 workloads and evaluate how runtime preparationaffects the performance. The experiment is performed withstable LAN WiFi, in which we suppose the best networkenvironment is provided and take no account of the instabilityof network.We divide the process of offloading into 4 phases. Computation Execution is the pure execution stage of offloadedtasks on the cloud. Runtime Preparation is the setup phase formobile code runtime after the offloading requests arrive. Network Connection is the process of establishing a connectionbetween mobile devices and cloud resources. Data Transferis the time spent to transfer necessary data for offloadingtasks. Figure 1 shows phase details and offloading speedups ofB. OS-level VirtualizationWhile the traditional system virtualization provides strongfault isolation, OS-level virtualization is much more efficient,especially for server consolidation. Containers, as the center of OS-level virtualization, have recently emerged as alightweight alternative to hypervisor-based virtualization [15].Unlike virtual machines, containers usually imposes muchless overhead, since they share the host OS kernel and donot suffer the overhead of resource virtualization.Unfortunately, the requirement of sharing host OS kernelmakes OS-level virtualization unsuitable for mobile cloud, asgeneral purpose OSes are incompatible with the mobile OSand does not support running mobile code. This limitation isa vital challenge of designing a container-based platform formobile cloud computing.1 http://www.computerchess.org.uk/ccrl/4040/124