Details for Talk on: 28.11.2017

  • Speaker: Ludwig Dierks

  • Title: Cluster Pre-scheduling: Combining Machine learning and market design 

  • Abstract: Cloud computing centers continuously run ten thousands of deployments, each consisting of one or more virtual machines (VM's). Here each center is divided into different clusters of physical hardware and arriving deployment requests get sent into different clusters by a central controller. The cluster now needs to decide whether to accept the new deployment for scheduling. Usually this is done by a fixed threshold  on the current utilization. But as in modern cloud systems, existing deployments can scale out and request more VM's on the same cluster, this threshold has to be set very low. In this talk we explore an approach that combines learned prior information on deployments with incentives for users to keep their deployments predictable in order to derive a better acceptance policy based on the first two moments of deployments.