../tc-2023-08-0440

Review: TC-2023-08-0440 Effective Huge Page Strategies for TLB Miss Reduction in Nested Virtualization

Paper summary

This paper aims to address the high TLB miss rate and slow address translation problem caused by discrepancies in page sizes and misalignment across host and guests under nested virtualization environment.

Problem definition

Discrepancies in page sizes and misalignment across host and guests diminish the effectness of huge pages in reducing TLB misses and speedup address translation.

Contribution

Comments

Abstract

Introduction

Nested Virtualization (section 4.2)

Three main challenges

  • Interposed hypervisor interfer with host/guest page align coordination
    • In addition, changing to the Interposed hypervisor might not be viable
  • Interposed hypervisor use base pages while both guest and host use hugepage
    • i.e. 1x guest hugepage -> 512x hypervisor basepage -> 1x host hugepage
  • Application alignment mismatch

Design (section 5)

gemini xgemini

Major components:

  1. HA: huge aligner. Allocate the entire hugepage instead a base page when a virtual address is touched.
  2. NEMA: nested enhanced memory allocator. Ensure GVA/GPA/HPA alignend to the same offset after hugepage boundaries when a guest page fault occurs. This ensures only one TLB entry can be shared by the entire hugepage the faulting address in.
  3. HB: huge booking
  4. MHPP: misaligned huge page promoter
  5. NMHPS: nested misaligned huge page scanner

Evaluation (section 6)

Objectives:

  1. Improve throughput-oriented applicaiton performance (PARSEC suite)
  2. Improve mean and tail latency of latency-sensitive workloads (Tailbench suite)
  3. Applicability and overhead (Xapian, Ferret)

Misc questions