• No results found

From these results we conclude that allocation schemes generated when MGECP is the prime objective with MMUP as a by objective have good level of optimality in achieving both objectives in MGECP and MMUP. Solving MGECP not only reduces the energy cost of the MCP as a whole, but also in some degree reduces the maximum utilisation of the MCP and boost the lifetime of the MCP as a platform. When MMUP is a priority, the SA algorithm can be used to further reduce the maximum utilisation of the MCP.

3.6

Summary

In this chapter, we first introduced the common structure of a multi-device and multi-workflow mobile cloud platform and the task scheduling problems in such platforms. This distinguishes our work from existing researches which model only the task scheduling problem in single-device and single-workflow scenarios of mobile cloud computing.

Two aspects of the energy-aware requirements of a mobile cloud platform are investigated in this chapter, namely the Minimum Group Energy Cost Problem (MGECP) and the Minimum Maximum Utilisation Problem (MMUP). In MGECP, our objective is to minimise the overall energy cost of the platform, whereas in MMUP, our objective is to maximise the energy-life of the platform. Both problems are realistic and critical to improving the energy-efficiency of mobile cloud platforms.

In order to model both objectives of the energy-aware task scheduling problem of a mobile cloud platform, we first characterised the computation and communication costs of the platform by abstracting the key parameters of the platform. We then summarised the total energy cost of a mobile cloud platform with a quadratic binary program the solution of which provides an optimal allocation scheme of the MGECP. Next, a quadratically constrained variant of the quadratic program was developed to model the the MMUP.

Due to the computational complexity of solving the quadratic programs to optimal, we devel- oped two heuristics to approximate the optimal solution for both scheduling problems in MGECP and MMUP. In the first heuristic, we implemented a simulated annealing algorithm (SA) which is often applied to combinatorial optimisation problems, especially for quadratic programs, in literature. In the second heuristic, we took a greedy (GAO) approach to the problems, which

allows each device to make task offload decisions autonomously. We tailored both algorithms to fit the requirements of our problems so that allocation schemes can be produced efficiently.

A comprehensive simulation study is carried out to verify and compare the performance and quality of our algorithms to that of the solvers provided by CPLEX, an industry-leading opti- misation package. In these simulations, we also applied two variants of our standard simulated annealing algorithm to demonstrate and verify our choice of parameters in the standard simulated annealing algorithm. Furthermore, we implemented two joint search algorithms which combines SA and GAO in an attempt to overcome the disadvantages of each of the two heuristics.

From the simulation results, we evaluate each algorithm from two aspects: the optimality of the allocation schemes produce by the algorithm, and the amount of time it took the algorithm to execute. For the MGECP, we find that GAO is the best algorithm amongst all in terms of both solution optimality and execution time. The joint search method SA+GAO is able to improve the optimality of the solution while costing more time to execute. While guaranteeing the optimal scheduling schemes, the QP solver from CPLEX is only applicable for small scale problems for its high and unpredictable execution time. For the MMUP, we find that SA produces the best solution, while GAO’s time cost remains small. The QCP solver from CPLEX is not able to produce good solutions within five times the execution time of the standard version of our SA algorithm.

Finally, we compare our results from MGECP and MMUP in terms of their contribution to each other’s objective. Although both problems are critical in an energy-aware mobile cloud platform, we wish to see the effect of solving one of the problems on the other, so that a decision can be made when only one problem (objective) is to be selected. We find that solving MGECP also has a positive effect towards our goal in the MMUP. On the contrary, solving MMUP alone increases the overall energy cost of the MCP which is opposite to the goal of MGECP.

Based on our findings in this chapter, we see that a tailored greedy algorithm is able to efficiently produce good solutions for energy-aware task scheduling problems in mobile cloud computing scenarios.

The modelling technique we presented in this chapter is also applicable to other energy critical scenarios.

Chapter 4

Offloading Strategies for

Time-Constrained Mobile

Workflows

Whereas energy cost is the priority for mobile cloud platforms, time cost is of equal importance when an individual workflow is concerned. As well as energy, time is another important metric associated with mobile application workflows. In this chapter, we investigate further into the task allocation problem within a mobile cloud environment and look at developing offload strate- gies for mobile workflows while taking into account both time and energy requirements of the workflow.This distinguishes our work to existing researches which consider only one of these two aspects (in time efficiency [30, 34, 38, 74] and in energy saving [27, 28]). A similar analysis on the offload-abilities of tasks is included in [75], but not in any great detail, and also is only based on single smartphone nodes.

To broaden our vision of a ubiquitous mobile cloud environment, we further introduce the concept of cloudlets into our platform model which acts as an additional layer of execution platform in our mobile cloud platform. Cloudlets as proposed in [38] are not as powerful as standard cloud services. The advantage of cloudlets is their accessibility to mobile devices.

Located at the edges of the network, they have very close physical proximity to the mobile devices. This greatly reduces the communication energy cost to the mobile cloud platform and time cost to the mobile workflows.

Also different from the scenarios we looked at in Chapter 3 is that, in this chapter, we assume all tasks are originally located on mobile devices. When an offload action is scheduled, the executables associated with that task must be transmitted to the cloud or cloudlet before that task can be executed remotely. This creates extra communications between mobile and cloud. Consequently, this extra cost need to be considered when offloading decisions are made.

We develop an algorithm WGAO to produce the offload strategies in this chapter which is based on the GAO algorithm. A comprehensive analysis of the simulation results give further insight of the relation between different characteristics of a workflow and its offload-ability.

Related documents