Mobile Offloading. Matti Kemppainen

Transcription

1 Mobile Offloading Matti Kemppainen

2 1. Promises and Theory Learning points What is mobile offloading? What does offloading promise? How does offloading differ from earlier practices? What is needed to implement an offloadingenabled environment?

3 Problem Description Slide by Prof. Jukka K. Nurminen Energy Consumption Battery Capacity Mobile Phone Usage Web Social media Multimedia Navigation New Services Basic Use (calls, SMS) less new services more frequent battery charging physically larger battery more energy-efficient components a breaktrough in battery cell technology a more clever way to utilize the available power

4 Solution? Mobile Computation Offloading transfer of computation execution outside the mobile device results Surrogates computation

5 Some Terminology Client Local Mobile Node Server Remote Cloud Augmenter Computation Node results = state + metadata Surrogates computation = state + metadata

6 Gaining Benefit Technical dimension Offloading is beneficial, if the related overhead costs are less than the cost of computation done locally.

7 Some Application Examples Categorization adapted from Chun & Maniatis Primary functionalities speech, video processing! Background tasks web crawling, photo analysis! Hardware augmentation speedup with more resources, specialized resources! Multiple execution paths artificial intelligence, different analysis methods!

8 Motivation!!Constraints Saving Energy Enhancing Reliability Enabling Performance Exploiting Context Monetary Cost Security and Trust Code Migratability, Limits of Automation Easiness for Application Developers

9 Key Features of Offloading Frameworks Migration Support The system that lies under the application code supports remote migration inherently. no need for application-specific networking protocols Offloading As an Alternative Remote execution is an opportunistic alternative, not a must. Offloading is an optimization method, not a requirement. Dynamic Decisionmaking Environmental conditions (e.g. state of network access, battery level) may have an effect on the execution location.

10 Software Architectures Levels of Offloading Feature idea: implement features and use them through an interface example: a typical network-enabled mobile application Method idea: execute resource-hungry methods remotely example: AI analysis of game logic System idea: clone the runtime environment (or the relevant parts) example: everything that might run on a system

11 Software Architectures Feature Offloading offload sematically coherent parts of the application Cuckoo (Android) Vrije Universiteit, Amsterdam Requirements: Standard Dalvik VM and Android software stack 1. Developer defines an interface (in AIDL) for the part of the application that is subject to offloading. 2. Building system generates the needed implementation stubs and proxies. 3. Developer implements the features. Local and remote implementations may differ.

12 Software Architectures Method Offloading offload method calls including needed data Example: MAUI (.NET) Duke, U. Mass. Amherst, UCLA, Microsoft Research Requirements: Standard.NET software stack 1. Developer annotates the desired methods as remoteable. 2. Framework considers offloading of the remoteable methods. It may also choose to invoke a method locally.

13 Software Architectures Image Offloading offload bytecode, program image or even volume image CloneCloud (Android) Intel Labs, Berkeley Requirements: A custom version of Dalvik VM 1. Developer lets the underlying system make partitioning and offloading decisions.

14 Architecture Comparison In theory! Developer workload Level of automation Needs platform support Feature Method System high medium low low medium high no not necessarily yes

15 Decisionmaking Mobile environment is inherently dynamic. Therefore we need dynamic decisions. Prior Analyses Developer s Decisions Application Profiling CPU usage, memory consumption network usage disk I/O Use-case Profiling Runtime Analyses User s decisions Environment Profiling hardware resources, network availability! Action Monitoring feedback-driven controlling of offloading process

16 Process State Computation is done on the data. The data needs to be available for the party that executes the computation. Costly Data Transfer transfer as little as possible Serializability data needs to be transferrable hardware driver cannot be offloaded class inheritance may pose considerable problems Complexity of Automation What is the needed dataset?

17 Networking The major difference of offloading and traditional distributed computing (e.g. grid) is in the networking. Mobility! wirelessness Sparse connectivity Multitude and heterogeneity of network stacks Energy consumption of antenna amplifier Long RTTs, packet loss Some Resolutions Network stack abstractions Traffic shaping Route selection (data offloading)

18 Infrastructures Runtime environment for the migrated code different implementations or a common software stack? Cloud services virtualization as a way to providing a suitable environment Networking performance surrogates closer to the clients Existing resources private clouds, PCs, specialized processors, other network devices in the local environment

19 Other Considerations Code Transfer application surrogate application libraries Data Transfer Optimizations transfer deltas delay tolerance of data Service Discovery mainstream users don t want to configure IP addresses Trust And Security how to make offloading trustable?

20 Some Cool Ideas Popular services brought nearby e.g. many subscribers for a newspaper on an aeroplane! a clever proxy that retrieves personalized content Universal Application Execution one application with two different interfaces for supporting many terminals (e.g. desktop computer, mobile phone) transfer of live process with help of offloading

21 2. Experiences from Practice

22 ThinkAir Deutsche Telekom (some modifications in Aalto) Method-level offloading framework runs on default Dalvik VM, no modifications needed the target application essentially an offloading client Any existing software requires modifications accordingly Client-server networking paradigm server application runs inside an unmodified Dalvik VM, too automatic application code transfer Designed to be an academic PoC tool building system requires a lot of manual operation code base suitable for measurements, not for real life use

23 Motivation Promises for what? MAUI: 45% energy savings for Chess AI CloneCloud: 20x speedup and energy savings for a large image search Biased measurements? Majority of earlier offloading measurements are conducted with a tailored application set Effect of networking energy consumption Could offloading be utilized for traffic shaping?

24 Testing Environment Google Nexus One, Google Nexus S mobile phones android-x86 running in VirtualBox Two network setups 1. Client and server in the same network 2. Client accessing server s network through 3G and Internet A few test applications Simple configuration tests and demos AndTweet Twitter client Static decisions dynamic decisions are not that well supported

25 Experiences And Results Success! We offloaded successfully ThinkAir handled the necessary procedures for execution migration Communication Offloading brought some advantage with WLAN!but! We did modify the application for custom serialization 3G RTT nullified the advantages of offloading Debugging is hard Errors may even be unnoticed Java is not an ideal environment for method-level offloading

26 Beneficial Offloading? There is no point in offloading if it is not beneficial. How can we identify a good use case for offloading? Application Developer trust on the developer s knowledge and intuition Use-case Analysis measurement of method-specific frequency and energy consumption

27 Code Migratability Given that we know what would be beneficial to offload, is it really offloadable? Migratability Criteria method migratability (simple version): does not access physical resources of the mobile device a small study showed that 85% of all methods in 16 different open-source applications were not migratable

28 Role of the Application Developer Application Developer must be an active part of offloading process!

29 3. Conclusions The key points once again!

30 Theory Mobile Computation Offloading Mobile Computation Offloading means means transfer of computation outside the mobile device. related terminology is emerging while research continues MCO differs from other distributed computing. opportunistic operation low-quality networking environments Offloading brings many potential benefits. energy saving, performance, reliability, ease for the software developers, better exploitation of contextual information... Offloading has also many other opportunities. business opportunities, collaborative local services, universal application execution!

31 Practice Current State of Art Today s frameworks deal with the essentials. There is no publically available offloading framework. Current frameworks are more or less for academic purposes.

32 Future Prospects Ongoing research in Aalto DCS business feasibility analysis toolkit for offloading existing applications Ongoing research elsewhere (acc. papers) Cuckoo: dynamic decisionmaking! MAUI: state transfer optimization! CloneCloud: hardware accesses, advanced concurrency, trustability!