Scheduling Virtual Grids: The Magrathea System

Contemporary production grids do not usually offer the flexibility users are looking for. While different user communities have often contradictory requirements on the operating system, libraries, and applications, the production Grids provide only one rigid environment. This rigidness can be overcome by virtualization, when every user community or even individual user can be provided with its own instance of a virtual Grid, running optimized and tailored operating system and services. The promise of higher flexibility of virtual Grids is compensated by the increase in scheduling complexity. In this paper, we present the Magrathea system that extends the Grid resource management systems with support for virtual environment. After discussing the design requirements, we introduce the Magrathea architecture that consists of three components: the master and slave processes running on virtualized resources and the cache process to provide the information about virtual machine state to the scheduler. Two virtual machines sharing one physical resource and used exclusively, preemption of a lower priority job running in a virtual machine, support for more than two concurrently domains and support for "frozen" services that are repeatedly invoked and suspended are the use scenarios discussed in the second part of the paper. We demonstrate how they are supported by the Magrathea system and what modifications to the Grid resource management system are necessary. The Magrathea is currently in the pre-production use on the Czech national Grid environment META Center.