Supporting Deadline Constrained Distributed Computations on Grids

The growing popularity of grid and cloud computing has led to a renewed interest in resource control and coordination. The Actor model, which encapsulates objects along with threads of execution, offers a convenient way for scheduling computations' access to resources by way of scheduling of the actor threads. However, efficient Actor implementations do not use a thread for each actor, making implementation of fine-grained resource scheduling decisions difficult. This paper presents our work on integrating mechanisms for deadline assurance into an optimized implementation of Actors. We achieve this by using deadline-driven adaptive scheduling, which prioritizes individual message deliveries and method executions involved in a distributed computation, based on the calculated deadlines by which each must be completed. These deadlines can be efficiently calculated at run-time for an important class of computations which use pipeline interaction style. Additionally, a tuner dynamically balances -- manually or automatically -- the overhead of the control mechanisms against the extent of control exercised. Experimental evaluation shows that the approach offers effective support for timeliness requirements (for multimedia QoS, for example) at the cost of a relatively modest and adjustable overhead.