A Meta-scheduler with Auction Based Resource Allocation for Global Grids Saurabh Kumar Garg, Srikumar Venugopal and Rajkumar Buyya Grid Computing and Distributed Systems Laboratory Department of Computer Science and Software Engineering The University of Melbourne, Australia {sgarg, srikumar, raj}@csse.unimelb.edu.au Abstract As users increasingly require better quality of service from Grids, resource management and scheduling mecha- nisms have to evolve in order to satisfy competing demands on limited resources. Traditional schedulers for Grids are system centric and favour system performance over increas- ing user’s utility. On the other hand market oriented sched- ulers are price-based systems that favour users but are based solely on user valuations. This paper proposes a novel meta-scheduler that unifies the advantages of both the systems for benefiting both users and resources. In order to do that, we design a valuation metric for user’s applications and computational resources based on multi-criteria re- quirements of users and resource load. The meta-scheduler maps user applications to suitable distributed resources us- ing a Continuous Double Auction (CDA). Through simula- tion, we compare our scheduling mechanism against other common mechanisms used by current meta-schedulers. The results show that our meta-scheduler mechanism can satisfy more users than the others while still meeting traditional system-centric performance criteria such as average load and deadline of applications 1. Introduction Computational resources in large Grids are generally managed by meta-schedulers that interface with the lo- cal job schedulers at each resource such as Portable Batch Scheduler (PBS), Load Sharing Facility (LSF) and LoadLeveler, to determine the most appropriate resource for executing a job submitted by a Grid user. Examples of such meta-schedulers include Maui/Moab scheduling suite [1], Condor-G [2], gLite Workload Management System [3] and GridWay [4]. These meta-schedulers mostly focus on im- proving system-centric performance metrics such as utiliza- tion, average load and applications’s turnaround time [5]. While the Grids have become more mature with re- spect to the integration of different components, users have also developed more sophisticated requirements, and are ready to pay upto a certain limit to satisfy them. Current meta-schedulers are unable to satisfy such requirements as they do not consider users’ urgency and resource valuation. Thus, we need new scheduling mechanisms which are not only efficient but also that take into account user interests, resource valuation and demand; and schedule user applica- tion jobs in a fair manner [13]. In recent years, a number of researchers have proposed economy-based models for more efficient management of Grid resources [6][7][8]. Such models apply well-known and proven economic mechanisms such as markets and auctions to solve the challenges of resource allocations in shared distributed computing environments. Auctions have been particularly preferred by many such projects − for example, Tycoon [7] and Bellagio [8]– as they provide a decentralized structure, are easy to implement, provide immense flexibility to participants to specify their valua- tions and are considered as the most efficient among current market management systems [9][20]. But these economic- based systems have many limitations. First, while these ap- proaches distribute services fairly, they limit the ability of customers to express fine-grained preferences for services. In addition to that, users may not be able to express their true valuations accurately as they may lack the sophistica- tion to make decisions based on changing resource load and prices. Finally, users with low budgets and urgent require- ments may not be able to gain resource allocation as the system may be monopolized by those with large budgets. This paper presents an auction-based meta-scheduler that aims to overcome the afore-mentioned limitations by taking into account not only the user valuations and re- source prices but also other important factors such as re- source load and waiting time for the jobs. The meta- scheduler matches jobs to resources using Continuous Dou- ble Auction (CDA). The job and resource valuation are then dynamically change depending on the urgency of the job and the load on the resources. We evaluate this mechanism