Scribd
Upload
84 views

A Comparison of Software and Hardware Techniques For x86 Virtualization

This document compares software and hardware techniques for x86 virtualization. Classical virtualization methods like deprivileging and using primary/shadow structures were inefficient due to frequent traps. Software virtualization uses binary translation to avoid traps, though hardware virtualization extensions directly execute guest code. Experiments show that while hardware enables near-native speeds, software virtualization is preferable for I/O workloads and hardware for system calls. Overall, both software and hardware virtualization perform equally despite maturity differences.

Uploaded by

Ahmed Hamouda
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
84 views

A Comparison of Software and Hardware Techniques For x86 Virtualization

This document compares software and hardware techniques for x86 virtualization. Classical virtualization methods like deprivileging and using primary/shadow structures were inefficient due to frequent traps. Software virtualization uses binary translation to avoid traps, though hardware virtualization extensions directly execute guest code. Experiments show that while hardware enables near-native speeds, software virtualization is preferable for I/O workloads and hardware for system calls. Overall, both software and hardware virtualization perform equally despite maturity differences.

Uploaded by

Ahmed Hamouda
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 18

A Comparison of Software and

Hardware Techniques for x86


Virtualization

Keith Adams, Ole Agesen


VMware

Presented by Keith Rutkowski


Overview
● Introduction
● Classical Virtualization
● Software Virtualization
● Hardware Virtualization
● Comparisons / Experiments
● Conclusion
Introduction
● Historically x86 has lacked HW virtualization
● Paravirtualization and guest OS modding
produces great results, but not always possible
● Currently best VMMs use binary translation
● Now with AMD and Intel providing virtualization
extensions, we can evaluate performance
between hardware and software virtualization
Classical Virtualization
● What is “virtualization”?
– Popek & Goldberg in 1974
– Fidelity
● Software on VMM executes identically to on hardware
allowing for some timing effects
– Performance
● Majority of guest instructions executed directly by
hardware
– Safety
● The VMM manages all hardware resources
Methods of Classical Virtualization
● De-privileging
– Run guest OS at reduced privilege level so that
privileged instructions trap to the VMM
– VMM then emulates the HW performing the
instruction
● Primary and shadow structures
– Privileged state of HW is different than guest OS's
expectation
– Fake the privileged HW register values in a shadow
structure
Methods of Classical Virtualization
● Memory traces
– Needed to keep coherency of shadow structures
– Guest OS page table entries are write-protected so
shadow copies can be modified when trap occurs
– DMA devices must be protected from reads and
writes to trigger a trap for VMM to keep shadow
copy of memory region
● Traps are expensive, so what else can we try?
– Improve the guest OS / VMM interface
– Improve the VMM / hardware interface
Software Virtualization
● Interpreter
– Literally pretends to be an x86 processor executing
instructions
– BAD!!! One interpreted instruction requires
hundreds of real instructions to execute
● Binary translation
– Instructions are “interpreted” on demand in chunks
– Translated instructions are a safe instruction subset
– Most instructions are 1:1 translation
– Re-addressing concerns for accessing memory
Software Virtualization
● Binary translation
(source)
Software Virtualization
● Binary translation (translated)
Software Virtualization
● Adaptive binary translation
– Detect frequently trapping code, retranslate to avoid
traps, patch new code into execution chain
– Group load/stores in “batch” calls to trap only once
Hardware Virtualization
● Provide virtual registers to fool guest OS
● Introduce a guest privilege mode
– “vmrun” instruction switches to guest mode for
guest OS execution
– traps return back to normal kernel mode for VMM
execution
● Still no MMU virtualization though... =(
Comparisons / Experiments
● Software vs. hardware
– Trap elimination – adaptive BT replaces most traps
which hardware must encounter
– Emulation speed – adaptive BT predecoded
instructions beats hardware “accelerated” traps
– Callout avoidance – BT avoids callout costs, but
hardware performs them faster still
– Code density is preserved among the two
● Hardware and software suffer different
overheads and causes for the overheads
Comparisons / Experiments
● Mature SW vs. first generation HW
Comparisons / Experiments
● Mature SW vs. first generation HW
Comparisons / Experiments
● Mature SW vs. first generation HW
Conclusion
● New hardware extensions allow x86 direct
guest OS execution
● Hardware vs. software
– Both execute computational bound code at near
native speed
– Software VMM is preferable for I/O bound code
– Hardware VMM is preferable for system call bound
code
● Adding MMU virtualization benefits to HW will
help
Conclusion
● Hybrid solutions could offer the best of both
worlds
● Despite the maturity of the software solutions
and the recent appearance of the hardware
solutions, both perform equally
● Most practical scenarios benefit from software
virtualization though
Questions?

You might also like