1. Xen: The Art of Virtualization
Paul Barham , Boris Dragovic, Keir Fraser, Steven Hand, Tim Harris, Alex Ho, Rolf Neugebauer, Ian Pratt, Andrew Warfield
Presented By: Tim Trippel
EECS 582 – W16

2. Overview Design Goals Xen Architecture Subsystem Virtualization
Performance Evaluation

3. Design Goals
Challenge: no hardware support for virtualization at the time

4. Design Goals
Challenge: no hardware support for virtualization at the time
Performance: Run multiple (100) different OSes on single machine

5. Design Goals
Challenge: no hardware support for virtualization at the time
Performance: Run multiple (100) different OSes on single machine
Deployability: Support Unmodified ABIs

6. Design Goals
Challenge: no hardware support for virtualization at the time
Performance: Run multiple (100) different OSes on single machine
Deployability: Support Unmodified ABIs
Security: Strong hidden isolation

7. Hypervisor Architectures
Guest OS Processes

8. Xen Architecture

9. Xen Architecture
Hypercall

10. Xen Architecture
Async. Event

11. Full Virtualization vs. Paravirtualization
Full Virtualization = virtual hardware exposed is functionally identical to machine hardware
Paravirtualization = virtual hardware is similar to underlying machine architecture
Full Virtualization:
PRO: no need to modify guest OSes
CON: not all architectures were designed for full virtualization (i.e. x86)
more privilege levels needed
MMU tough to virtualize since hardware handles operations
situations when real and virtual resource visibility is desirable
timers (TCP timeouts)
Machine addresses (improve memory paging performance)
Paravirtualization:
PRO: improved performance (and easier to virtualize x86 architecture)
CON: must modify OS slightly

12. Full Virtualization vs. Paravirtualization
Full Virtualization = virtual hardware exposed is functionally identical to machine hardware
Paravirtualization =
Full Virtualization:
PRO: no need to modify guest OSes
CON: not all architectures were designed for full virtualization (i.e. x86)
more privilege levels needed
MMU tough to virtualize since hardware handles operations
situations when real and virtual resource visibility is desirable
timers (TCP timeouts)
Machine addresses (improve memory paging performance)
Paravirtualization:
PRO: improved performance (and easier to virtualize x86 architecture)
CON: must modify OS slightly
Requires Modification of Guest OS

13. Full Virtualization vs. Paravirtualization
Full Virtualization = virtual hardware exposed is functionally identical to machine hardware
Paravirtualization =
Full Virtualization:
PRO: no need to modify guest OSes
CON: not all architectures were designed for full virtualization (i.e. x86)
more privilege levels needed
MMU tough to virtualize since hardware handles operations
Paravirtualization:
improved performance (and easier to virtualize x86 architectures
must modify OS slightly
Xen

14. Virtualizing Memory Management
Recall virtual address translation:
Easier if TLB is software controlled
Easier if TLB is tagged

15. Virtualizing Memory Management
Recall virtual address translation:
Easier if TLB is software controlled
Easier if TLB is tagged by address-space
Flushing TLB is EXPENSIVE → especially for Shadow Paging
Both NOT supported by x86 at the time:
TLB Miss - PT must be HW accessible
TLB Flushes - Context Switches

16. Shadow vs. Direct Paging
Flushing TLB is EXPENSIVE → especially for Shadow Paging

17. Shadow vs. Direct Paging
Validate PT Update w/ Hypercall

18. Shadow vs. Direct Paging
Validate PT Update w/ Hypercall
READ ONLY!

19. Avoiding TLB Flushes
Allow Xen to exist in top 64MB of every address space (does not violate ABI)
Top 64MB is not remappable by guest OS
Xen Memory
Xen Memory
Context Switch
OS Memory
OS Memory
Process Memory
Process Memory

20. x86 Architecture at the Time
Virtualizing the CPU
Ring 3 - Applications
Ring 2 - Unused
Ring 1 - Unused
Ring Compropression
Ring 0 - OS Kernel
x86 Architecture at the Time

21. x86 Architecture at the Time
Virtualizing the CPU
Ring 3 - Applications
Ring 2 - Unused
Ring 1 - Unused
Ring Compropression
Hypervisor?
Ring 0 - OS Kernel
x86 Architecture at the Time

22. Virtualizing the CPU x86 Xen Solution Ring 3 - Applications
Ring 2 - Unused
Ring 1 - Guest OS
Ring Compropression
Ring 0 - Hypervisor
x86 Xen Solution

23. Separate Address Spaces
Virtualizing the CPU
Ring 3 - Applications
Separate Address Spaces
Ring 2 - Unused
Ring 1 - Guest OS
Ring 1 - Applications
Ring 1 - Guest OS
Ring Compropression
Ring 0 - Hypervisor
Ring 0 - Hypervisor
x86 Xen Solution
Alt. Xen Solution

24. Virtualizing the CPU x86 VMX Technology Ring 3 - Applications
Ring 2 - Unused
Ring 1 - Unused
x86 VMX Technology
Ring 0 - Guest OS
Ring Compropression
Ring -1 - Hypervisor

25. Virtualizing Hardware Devices
Do not fully virtualizing specific devices with emulation
Do create clean device abstraction
asynchronous descriptor rings

26. Virtualizing Hardware Devices
Asynchronous descriptor rings
allow Xen to reorder ops for performance
allow zero-copy transfer
VM can batch requests before invoking hypercall
VM instruct Xen to batch responses

27. Virtualizing Networking
DomX Virtual Network Interface
VIF = Virtual Network Interface
Two descriptor rings (RX & TX)
VFR = Virtual Firewall Router
Dom0 installs/maintains rules
Round-robin packet scheduler TX RX
Virtual Firewall Router

28. Virtualizing Scheduling & Timers
Domain scheduling can be adjusted/managed by Domain0
implemented Borrowed Virtual Time algorithm to minimize domain wakeup for an event; i.e. correctly estimate TCP RTT
can implement other scheduling algorithm
Three Timers:
Real Time - nanoseconds passed since machine boot
Virtual Time - only advanced when domain is executing
used by guest OS scheduler
Wall-clock Time
offset of real time

29. Performance Evaluation - Single Guest OS

30. Scalability

31. Conclusion
Paravirtualization (PV) can offer performance gains over full virtualization
PV requires modifying OS
Virtualizing microarchitectures without hardware support is challenging

32. Discussion Points
With hardware virtualization support, is paravirtualization still worthwhile?
How is Xen’s device I/O abstraction similar to Arrakis?
How do Xen VMs compare with Docker Containers?