Bin repacking scheduling in virtualized datacenters

TLDR: This work introduces the Bin Repacking Scheduling Problem, a problem to find a final packing and to schedule the transitions from a given initial packing, accordingly to new resource and placement requirements, while minimizing the average transition completion time.

Abstract: A datacenter can be viewed as a dynamic bin packing system where servers host applications with varying resource requirements and varying relative placement constraints When those needs are no longer satisfied, the system has to be reconfigured Virtualization allows to distribute applications into Virtual Machines (VMs) to ease their manipulation In particular, a VM can be freely migrated without disrupting its service, temporarily consuming resources both on its origin and destination We introduce the Bin Repacking Scheduling Problem in this context This problem is to find a final packing and to schedule the transitions from a given initial packing, accordingly to new resource and placement requirements, while minimizing the average transition completion time Our CP-based approach is implemented into Entropy, an autonomous VM manager which detects reconfiguration needs, generates and solves the CP model, then applies the computed decision CP provides the awaited flexibility to handle heterogeneous placement constraints and the ability to manage large datacenters with up to 2,000 servers and 10,000 VMs