When eXtended Para - Virtualization (XPV) Meets NUMA

Abstract

This paper addresses the problem of efficiently virtualizing NUMA architectures. The major challenge comes from the fact that the hypervisor regularly reconfigures the placement of a virtual machine (VM) over the NUMA topology. However, neither guest operating systems (OSes) nor system runtime libraries (e.g., HotSpot) are designed to consider NUMA topology changes at runtime, leading end user applications to unpredictable performance. This paper presents eXtended Para-Virtualization (XPV), a new principle to efficiently virtualize a NUMA architecture. XPV consists in revisiting the interface between the hypervisor and the guest OS, and between the guest OS and system runtime libraries (SRL) so that they can dynamically take into account NUMA topology changes. The paper presents a methodology for systematically adapting legacy hypervisors, OSes, and SRLs. We have applied our approach with less than 2k line of codes in two legacy hypervisors (Xen and KVM), two legacy

Virtualization Overhead of Multithreading in X86 State-of-the-Art & Remaining Challenges

“…Rather, it depends on the host system P (S v ). All popular virtualization platforms are known to struggle with NUMA locality [4], [47].…”

“…Two methods are common: passing through the hardware NUMA architecture to the VM, and using automated heuristical NUMA-optimization algorithms at VMM level. The former method is available in every major VMM [4]. The latter is integrated in most common VMM and OS schedulers.…”

“…However, these benefits come at a cost, mainly in the form of efficiency and performance losses. Identifying and mitigating said losses is a long-standing challenge for researchers at all system layers [2], [3], [4], [5], [6]. Thanks to these efforts, most virtualized workloads can presently achieve near-native performance [7].…”

Closer: A New Design Principle for the Privileged Virtual Machine OS

“…This is unfair for uVMs which execute less I/O operations, but the unfairness is limited to step r 1 and mitigated as follows. In the hypervisor, we monitor the I/O activity of each uVM 3 and the notification of a vCPU from the SC (step r 2 ) is done in proportion to the I/O activity of each uVM (the uVM with the highest I/O activity will receive more notifications on its processors, i.e. SC's vCPUs pinned on these processors will be more solicited).…”

“…Para-Virtualization (PV), introduced by Disco [1] and Xen [2], is a very popular virtualization approach which consists in making the guest OS aware of the fact that it runs in a virtualized environment. This approach has been proven to reduce virtualization overhead [3] and also mitigate the limitation of some virtualization hardware features (e.g. the inability to live migrate a VM with SR-IOV [4]- [6]).…”