Towards Eliminating Memory Virtualization Overhead

Abstract

Despite of continuous efforts on reducing virtualization overhead, memory virtualization overhead remains significant for certain applications and often the last piece standing. Hardware vendors even dedicate hardware resources to help minimize this overhead. Each of the two typical approaches, shadow paging (SP) and hardware assisted paging (HAP), has its own advantages and disadvantages. Hardware-dependent HAP eliminates most VM exits caused by page faults, but suffers from higher penalties of TLB misses. On the other hand, the software-only approach, SP, enjoys shorter page walk latencies while paying for the VM exits to maintain the consistency between the shadow page table and the guest page table. We observe that, although HAP and SP each holds its ground for a set of applications in a 32-bit virtual machine (VM), SP almost always performs on a par with or better than HAP in a 64-bit system, which steadily gains its popularity. This paper examines the root cause of this inconsistency through a series of experiments and shows that the major overhead of shadow paging in a 32-bit system can be substantially reduced using a customized memory allocator. We conclude that memory virtualization overhead can be minimized with software-only approaches and therefore hardware-assisted paging might no longer be necessary.