Competitive Concurrent Distributed Scheduling
share
We introduce a new scheduling problem in distributed computing that we call the DSMS problem. We are given a set of k ≥ 1 mobile identical servers that are initially hosted by some processors of the given network. Further, we are given a set of requests that are issued by the processors possibly at any time in an online fashion, and the servers must serve them. A request must be scheduled before the request is served. The delay for scheduling a request is the time taken since the request is issued until it is scheduled. The goal is to minimize the average delay. Navigating mobile servers in a large-scale distributed system needs a scalable location service. We devise the distributed GNN protocol, a novel linked-reversal-based protocol for the DSMS problem that works on overlay trees. We prove that GNN is a starvation-free protocol that correctly integrates locating the servers and synchronizing the concurrent access to the servers despite asynchrony. Further, we analyze the GNN protocol for one-shot executions where all requests are simultaneously issued. We show that when running the GNN protocol on top of a special family of tree topologies---known as hierarchically well-separated trees (HSTs)---we obtain a randomized distributed protocol with an expected competitive ratio of O( n) on general network topologies for the DSMS problem where n is the number of processors in the given network. From a technical point of view, the main result of our paper shows that the GNN protocol optimally solves the DSMS problem on HSTs for one-shot executions. Our results hold even if communication is asynchronous.
READ FULL TEXT