Advancing on our previous work (contianer-based server runtime Rattrap), we implement an unikernel-based runtime for mobile computation offloading under Mobile Fog Computing or Mobile Edge Computing scenarios, Introducing much less boot-up delay, memory footprint, disk usage and energy consumption at IoT Edge. We firstly put forward the concept of Rich-Unikernel which aims to support various applications in one unikernel while avoiding their time-consuming recompilation. Following the design of Rich-Unikernel, we implement a not only lightweight but also flexible runtime for offloaded codes, called Android Unikernel, by integrating basic Android system libraries into OSv unikernel.
An offloading framework, which provides APIs for mobile applications to quickly offload local computation-intensive tasks to server side to execute. See examples in Applications
A scheduling-programe——Dispatcher. It is in charge of handling offloading requests and booting up an unikernel server for each request.
This part is to build into unikernel. It contains a tool called DynamicLinker which is able to load class from Android bytecodes(.dex) and link application into unikernel at runtime. When an unikernel boots up, this programme runs automatically and waits for offloading requests. With the help of DynamicLinker, we can omit a lot of time-consuming recompiling work when building application into its corresponding unikernel.
This part contains common Android system libraries that may be used by offloaded codes. It is also packaged into unikernel image along with DynamicLinker at compile-time.
Since traditional unikernel takes a few seconds to be specicalized for an application and needs to be reconstructed when application changes even if the change is very small, it is not suitable for offloading scenarios which often meet various applications and need to response to requests timely. Therefore, We define Rich-Unikernel, a kind of unikernel that is more general than conventional unikernel. Rich-Unikernel is not specialized for a particular application and it can be regarded as a base unikernel for a series of applications. All of the system libraries needed by these applications have already been packaged into the base unikernel, so it is able to run different applications (one application at a time).
In our system, we implement a kind of Rich-Unikernel called Android Unikernel, by integrating common Android libraries that are needed by offloaded codes into OSv unikernel. Also, We impletement a tool called DynamicLinker, which is able to dynamically link application codes into Android Unikernel at runtime. Thus, it can serve various applications without time-consuming recompilation, greately improving the response speed of our system.
To build Android Unikernel, you needs the following environments:
- Install OSv (see https://github.com/cloudius-systems/osv/blob/master/README.md)
- JDK7
- Install KVM, libvirt
Following the example in which we provide some simple configureation files and scripts to build DynamicLinker and Libdroid into Android Unikernel.
Before testing, you should install MySQL and import unikernel.sql which saves the name, ip and status of a unikernle-based kvm. Dispatcher can quickly find a available kvm ( of which the status is 0) and start it when an offloading request arrives.
java -jar Unikernel-Scheduler
Download an app from Applications.
Here we present Linkpack.
Also, you can choose your own app and modify it like Linpakc does.
Start application and click setting to choose local execution or remote execution.
Local execution shows the result of running linpack on mobile devices itself. It takes 160 seconds to finish solving a set of linear equations.
Remote execution shows the result of running linpack through offloading. It takes only 58 seconds to solve the same linear equations.
