Scalevisor, a distributed hypervisor for rack-scale computing
Data-centers are under-utilized because they are struggling to consolidate virtual machines i.e. run a number of virtual machines on a fewer number of servers by occupying unused resources like memory, networking devices, compute units and storage solutions. Moreover, fragmentation of physical resources due to previous allocation and deallocation leads to degradation in resource utilization. Solving this problem with reasonable performances could reduce datacenter energy consumption but this problem is a bin-packing problem which is NP-hard thus no efficient solution can be found with the current architecture of hypervisors.
Using a distributed hypervisor would allow to allocate more space for one type of resource dynamically to tackle the problem of one particular resource shortage. This could avoid the need for consolidating VMs by expanding available resources. But distributed hypervisors were previously limited by the poor performances of interconnect compared to local buses. Recently, server vendors promises of "disaggregated hardware" for rack-scale computing should provide better performances for distributed hypervisor to run VMs with traditional operating systems.
My presentation at GDR RSD and ASF Winter School on Distributed Systems and Networks, 4th January, 2020.
My presentation at Journées Cloud 2020, 24th November, 2020.