提升KVM效能監測單位虛擬化之效能

Abstract

當今的處理器會透過特殊的性能分析硬體將架構相關的事件的資訊提供給使用者。具體而言，大部分的處理器架構都有提供效能監測單位 (PMU) 來記錄並且回報硬體事件的資訊。此外，有些供應商亦會在處理器上提供其他專門的性能分析功能，例如 Intel 處理器提供的最後分支紀錄 (LBR) 能夠提供最近執行的分支指令的詳細資訊。軟體開發者可以利用這些處理器提供的功能來對自己的程式做性能分析，並且根據分析的結果來優化程式。由於開發者越來越傾向於把程式部署於雲端服務所提供的虛擬機器，主流的虛擬機器監測器會虛擬化這些性能分析硬體，並且將虛擬的性能分析硬體提供給虛擬機器使用。然而，目前針對性能分析硬體的虛擬化會導致虛擬機器頻繁的退出到虛擬機器監測器，這也大幅提升了在虛擬機器中使用性能分析硬體的開銷。因此，這篇論文旨在優化 Intel 處理器的性能分析硬體的虛擬化的性能。我們在確保性能分析硬體被正確地多工複用的前提下虛擬機器直接存取性能分析硬體。此外，我們也把性能分析硬體產生的中斷直接傳遞給虛擬機器以降低頻繁取樣下造成的陷入。我們對我們在 Linux KVM 虛擬機器監測器上的實作原型進行評估，結果顯示我們能夠顯著地提高性能分析硬體如 PMU 的虛擬化效能。

Modern processors expose architectural events to users via special profiling features. Specifically, most architectures provide a performance monitoring unit (PMU) to record and report hardware events. Besides, vendors like Intel support customized features like Last Branch Record (LBR), which profiles the execution of branch instructions. Developers utilize these features to profile their programs’ execution and analyze performance. As programs are increasingly deployed to virtual machines (VMs) running on the cloud, commodity hypervisors expose virtual profiling hardware to VMs. However, the current virtualization support incurs frequent VM exits to the hypervisor, causing significant overhead to VMs using the profiling hardware. This thesis aims to optimize the virtualization performance of hardware profiling features on Intel processors. We multiplex the profiling hardware to safely enable direct access from the VM. Further, we pass through interrupts generated by the profiling hardware to VMs to avoid traps caused by frequent samplings. Evaluation of our prototype for the Linux KVM hypervisor shows that our approach significantly improved the virtualization performance of PMU.