Job migrations in queueing networks: some non-conventional product-forms

Speaker : Andrea Marin
University of Venice (IT)
Date: 25/09/2024
Time: 2:00 pm - 3:00 pm
Location: Amphi 6

Abstract

Job migrations are important to improve the performance in distributed systems.  In this talk, we explore two policies for job migration that are receiver-initiated, i.e., the migration of one or more jobs is triggered by the state of the stations that will receive the new workload. The first policy is aimed at improving the load-balancing in the queueing network, and allows an empty station to ‘steal’ a geometrically distributed batch of jobs from another station, if these are present. The second policy that we propose forces an empty station to steal one job from another station but, if this is empty, the job-stealing signal is propagated backward along the routing topology until a station with at least one job is found. We show that the queueing network models underlying these policies admit product-form solution under the usual exponential assumptions and discuss their mean value analysis.