Skip to content

Latest commit

 

History

History
61 lines (44 loc) · 2.72 KB

2020-11-22-Note-Sean2012An.md

File metadata and controls

61 lines (44 loc) · 2.72 KB
Raw
layout title date categories tags
post
Note: An Empirical Study of Memory Sharing in Virtual Machines.
2020-10-28 05:00:00 -0700
PaperReading
FileSystem
Deduplication

Reference

Sean Barker, Timothy Wood, Prashant Shenoy, Ramesh Sitaraman. An Empirical Study of Memory Sharing in Virtual Machines. In Proc. of USENIX ATC 2012.

What

Analyze and summarize the existing page sharing system, shows that real-world sharing is around 15% (less than expected) and studied a variety of factors impacting sharing (e.g., homogeneity)

Background

Page Sharing:

  • Eliminate identical pages of memory.
  • Virtual pages mapped to physical pages.
  • Hypervisor detects duplicates.
  • Replaced with copy-on-write references.

Page Sharing Types:

  • Self-Sharing: sharing within individual VMs.
  • Inter-VM sharing: sharing across multiple VMs.

Memory Trace generates:

  • Produce a memory trace from a VM by suspending the VM, then reading the resulting binary memory image to create the hash list

Paper Summary

Current page sharing systems remain questions like: (i). What levels of sharing are possible in typical real-world machines? (ii) What are the factors that impact sharing potential? (iii) How emerge in technologies impact sharing? This paper provides practical insights into these questions through a careful study of memory data with their own memory capture tools. It shows that about 15% of pages are possible sharing, while almost 85% sharing is derived from self-sharing. The major pages shared is come from libraries and heaps, which are possibly from recreated data structures. The factors impact page sharing is: (i) Hierarchy (OS family, application setup, OS version, OS architecture). (ii) Platform homogeneity: Minimal sharing across heterogeneous systems; Significant gains at homogeneous deployments. (iii) Finer-grained sharing may be leveraged to improve sharing potential. (iv) OS improvements like ASLR may reduce sharing

This paper shows that most page sharing not required virtualization and the sharing is around 15%, which is significant but less than expected.

Strength

  • For the first time, a detailed analysis of the value, source, and influencing factors of page sharing.
  • Build an in kernel page analysis tools that could capture the content type of each page and which process is using this page.

Weakness

  • Insufficient details of the experiment, such as the time interval and frequency of fetching the contents of the memory, are not described

Comments

  • Without open-source tools, many details in the experiment are difficult to reproduce.
  • If only 15% of the memory is shared, how effective is the actual resource sharing?