Comparisons of Memory Virtualization Solutions for Architectures with Software-Managed TLBs

Abstract

Memory virtualization plays an important role in system virtualization. However, traditional memory virtualization solutions are usually for those architectures with hardware-managed TLBs, such as x86, ARM and so on. The solutions for architectures with software-managed TLBs are rarely mentioned. This paper presents three different memory virtualization solutions for architectures with software-managed TLBs. The difference is that different methods are used to tackle guest TLBs. One is emulating the guest TLBs by software, another is eliminating the guest TLBs and the third is emulating guest TLBs by hardware. After comparisons on hardware platform with MIPS CPUs, it can be seen that the method eliminating guest TLBs can increase the memory bandwidth by more than 100% on average compared to the method emulating guest TLBs by software. After comparisons on simulation platform with MIPS CPUs, it can be seen that the method emulating guest TLBs by hardware would cause about 135% more extra exceptions than the method eliminating guest TLBs. The method eliminating guest TLBs is a good choice when full virtualization is not necessary on the CPUs with software-managed TLBs.