Evaluation of scheduling policies in a Mobile Grid architecture Konstantinos Katsaros and George C. Polyzos Department of Computer Science Athens University of Economics and Business Athens 104 34, Greece ntinos@aueb.gr, polyzos@aueb.gr Abstract— Recent advances in mobile communications and computing and strong interest of the scientific community in the Grid have led to research into the Mobile Grid. Based on a realistic Mobile Grid architecture we formulate the problem of job scheduling in a mobile environment and investigate key design decisions with respect to fundamental performance metrics. We extend previous work by introducing a new scheduling policy based on the notion of installments and continue the evaluation of the expanded set of scheduling strategies in an effort to overcome the intermittent character of connectivity in a mobile environ- ment. Our findings, based on real wireless traces, demonstrate the superiority of the proposed policy, show the feasibility of a Mobile Grid system and further provide valuable input for the design of efficient scheduling policies subject to the underlying mobility characteristics of the mobile resources. Index Terms— mobile grid, replication, intermittent connectiv- ity, performance evaluation, wireless traces. I. I NTRODUCTION The great advantages in mobile and wireless communica- tions have resulted in the availability of an enormous number of mobile computing devices such as laptop PCs and PDAs. In effect, it has been considered natural to extend the idea of resource sharing to mobile and wireless communication environments leading to the emergence of the Mobile Grid computing paradigm. However, there are various, quite dif- ferent approaches on the exact character of this extension. Their differences mostly rely on whether mobile devices are considered powerful enough to provide their resources or not. Some researchers argue that exactly because mobile devices are resource constrained, they should be incorporated in a grid only as resource consumers [1], [2]. Others support that mobile devices are increasingly becoming powerful enough (mostly in terms of computational power and storage) to also participate in grid systems as resource providers [3], [4], [5]. Mobile devices indeed face resource limitations, at least in comparison to their stationary counterparts, but the vast number of available mobile devices and the fact that their computational power is constantly increasing, lead us to the assessment that the aggregated sum of their resources could be exploited in order to overcome exactly these limitations. Towards this direction, we have proposed a hierarchical, campus-wide computational Mobile Grid system architecture [6]. In the proposed architecture, depicted in Fig. 1, mobile nodes (MNs), willing to offer their computational resources, move between Access Points (APs) of the campus. This willingness is based on the expectation of reciprocity [7]. In our work we have considered divisible load applications [8] (e.g. query processing [9]) in which a job can be divided into tasks that can be carried out independently of each other. These tasks are distributed by the Local-Mobile Grid Schedulers (L-MGSs) to the collaborating MNs, which process them and return the results back. In other levels of the hierarchy Intermediate-MGSs (I-MGSs) may act as meta-schedulers. This work focuses on the last level of the hierarchy and more specifically on the load distribution performed by L-MGSs. In previous work, we investigated the impact of intermittent connectivity on the performance of the proposed architecture. In our research we utilized the WLAN traces available at [10]. These traces provided us with realistic information on the mobility and connectivity characteristics of the MNs in the campus. We pointed out the important mobile networking parameters affecting the performance of a Mobile Grid system and showed that disconnection events impose a severe impact on the turn-around time of jobs executed by MNs. On an effort to smooth the effects of intermittent connectivity we examined the performance of a simple task replication scheme [6]. The results were very promising with respect to the achieved turn-around time. We continued our research by also investigating whether the execution of a task in a MN should be aborted upon disconnection or not, in order for the task to be rescheduled, coming up with a positive answer with respect to the resulting turn-around times [7]. In this work, we enrich the set of examined policies by proposing the installments policy, in which task load is further partitioned into small chunks, and demonstrate its effective- ness. We continue our investigation on the performance of each policy by also examining another important parameter, that of resource waste i.e. the amount of resources wasted in an effort to conceal intermittent connectivity. Moreover, we provide an analytical framework for the considered strategies that directly connects their suitability with the underlying net- working conditions. We apply this framework on our realistic, wireless trace environment and demonstrate its validity. The paper is organized as follows. In Section II we formu- late the problem of scheduling in the considered context and in Section III we describe the set of scheduling policies under evaluation. We further provide an analytical framework for the impact of intermittent connectivity on the performance of the