基於Arm平台Linux KVM虛擬機器之系統狀態回復

Abstract

近年來，在Arm平台上運行虛擬機器的應用越來越普遍。Google和Amazon等雲端服務供應商都提供Arm架構的虛擬機器。此外，新一代的車用系統，例如SOAFEE的硬體也使用Arm平台，並將車用軟體運行於虛擬機器中。車用軟體需要高故障容許度，以致在故障發生時能迅速恢復先前正常運行時的系統狀態，而記錄回復(Record Replay)便是提供故障容許度的機制之一。Record Replay在系統運行時紀錄系統運行狀態，並在回復時重現先前的系統運行情況。先前曾經有在其他架構及虛擬機器監視器的Record Replay之研究，但於Arm架構上之Linux KVM的Record Replay卻少有涉及。本論文討論在Arm平台上之Linux 基於核心的虛擬機器(Kernel-based Virtual Machine)的Record Replay設計，並在Linux v5.15及QEMU 7.1.0上實作了該設計的原型。實驗結果顯示，該原型的性能損失與原版KVM並不大，並且與虛擬機器快照(Virtual Machine Snapshot)的方法相比，本原型的性能損失較小。

In recent years, there has been an increase in the number of applications that run virtual machines on Arm-based platforms. Cloud providers such as Google and Amazon offer virtual machines that run on Arm-based processors. Additionally, new software architecture for vehicles, such as SOAFEE, also utilizes Arm-based processors with software running on virtual machines. Automotive software requires high fault tolerance to restore to its previous state in the event of software crashes or freezes. Record Replay is one such mechanism that provides fault tolerance by logging the system execution during the record stage and reproducing it using the logged information during the replay stage. Although research on Record Replay of virtual machines exists for various hypervisor and processor architectures, there is limited focus on the Linux KVM for Arm architecture. The thesis discusses the design of Record Replay of virtual machines on Linux KVM for Arm and presents the implementation of the prototype that builds on Linux v5.15 and QEMU 7.1.0. The experiment results indicate that the prototype has minimal performance overhead compared to the Vanilla KVM and has lower performance overhead compared to the Virtual Machine Snapshot method.