1. SAC 2011 -- OS track
Optimizing Virtual Machines Using
Hybrid Virtualization
Qian Lin
Shanghai Jiao Tong University

2. Agenda
Background: system virtualization
Motivation
Design and optimization
Evaluation
Summary

3. Background: system virtualization
Goal: consolidate and maximize H/W platform resources
Way: multiple VMs run concurrently on a single machine

4. Motivation
Current virtualization types
Software-only full virtualization
 Dynamic binary translation technique
 For cross-platform development and debug only
H/W assisted full virtualization
 Leverage H/W virtualization extension of CPU architecture
 Superior in CPU and memory virtualization, but not in I/O
Paravirtualization
 Ring compression → compromised strategy
 Exist much overhead in the execution of system call
 Efficient in I/O event handling

5. Motivation (cont.)
Reduce overhead incurred by virtualization
Optimize to reduce the execution redundancy
Hybrid approach to merge advantages
64-bit HVM is at par or faster than 64-bit PVM
Paravirtualization is helpful for enhancing HVM
 Simplicity, performance, efficiency, scalability, correctness

6. Design
Type I: Hybrid PVM Type II: Hybrid HVM
Port paravirtualized OS Local optimize HVM
Common design goal:
into HVM container using PV strategy
Run paravirtualized OS in the HVM container.
Add HAP support Import PV components

7. VM optimization (1)
Abridge the system call path
PVM: Kernel/User space share the same ring
Hybrid VM: Rings are assigned normally

8. VM optimization (2)
Shadow page table → complex and dull

9. VM optimization (2)
H/W assisted paging → accelerated!

10. VM optimization (3)
Local APIC is not used by Hybrid VM Linux
EOI (End of Interrupt) does not cause VM exit.
Use Xen API (event channel)
MSI/MSI-X handling is paravirtualized in Hybrid VM
MSI Mask/Unmask does not cause VM exit
No changes are made to device drivers
I/O intensive loads expose those virtualization overheads
About 12K interrupts/per sec. (per VCPU) with 10 GbE
CPU utilization is >3% higher per VCPU on ordinary HVM
Linux
As number of VCPUs increases, overhead increases

11. Micro performance (1)
LMBench - Processor
Overhead contributed by fundamental operations
Factor: system call

12. Micro performance (2)
LMBench - Context switch
Inter processes, kernel/user mode, interrupt
Factor: address translation, TLB flush

13. Micro performance (3)
LMBench - Local communication latency
Response time of I/O request
Factor: address translation

14. Macro performance: CPU intensive
Kernel compile benchmark
Factor: system call, context switch, address translation

15. Macro performance: I/O intensive
CPU utilization within Ethernet workload
Factor: interrupt handling

16. Summary
Minimize virtualization overheads and utilize
new/advanced H/W features in VMs as much as
possible for cloud computing
Take both superiority of H/W assisted
virtualization and paravirtualization
Open source implementation of prototype
Also release to the Xen community

17. Thank you!
Trusted Computing Group @ SJTU
http://202.120.40.124/index.php/Project_TC