Politecnico di Torino

Dipartimento di Automatica ed Informatica

TORSEC Group

Performance of Xen’s Secured Virtual Networks

Emanuele Cesena <cesena@mat.uniroma3.it>Paolo Carlo Pomi <paolo.pomi@polito.it>Gianluca Ramunno <ramunno@polito.it>Davide Vernizzi <davide.vernizzi@polito.it>

Outline Introduction Experiments Model Security mechanism Conclusion

Introduction

Motivations Server consolidation

Planning

Model of virtual network Emulation Comparison

Virtualization “Technique for dividing the resources of a computer

into multiple execution environments called virtual machines (VMs)” (A. Singh)

Full virtualization Complete emulation of the underlying hardware Unmodified operating system in the VM

Paravirtualization VM needs a modified OS Best performance, close to native

Virtualization: XEN XEN is a free Virtual Machines Monitor (hypervisor)

x86, Intel Itanium, PowerPC platforms Paravirtualization, full virtualization (hw support) Very low overhead when paravirtualized: average

3-5%

Virtual machines Domain-0: privileged VM

Direct access to hardware Direct interface to the hypervisor

Guest domains

Virtual Network in XEN Network interfaces

Front-end within VM: eth0 Back-end in Domain-0: virtual interface (vif) Connection between netfront and netback

provided by the hypervisor

XEN hypervisor

Domain 0

vif1.0 vif2.0

Guest 1

eth0

Guest 2

eth0

Virtual Network in XEN Virtual Network

Domain-0 manages all the netbacks Bridge as “L2-switch”

XEN hypervisor

Domain 0

vif0.0 vif1.0 vif2.0

peth0

eth0

br0

Dom-0 Guest 1 Guest 2

physicalworld

switch

Virtual Network in XEN Example: Guest 1 sends a packet to Guest 2

packet created within Guest 1 stack copied from FE to BE via page flipping forwarded through the bridge copied from BE to FE, then received by Guest 2

we call this a virtual link

Domain 0

vif1.0 vif2.0

br0

Guest 1

eth0

Guest 2

eth0

Experiments

Experiments HP Compaq dc7700

Intel Core2 Duo 2.13 GHz RAM: 2GB XEN 3.0.4 Linux kernel 2.6.20

10 Virtual Machines (guests) RAM: 128 MB Linux kernel 2.6.20 minimal Debian installation IPerf to test network bandwidth

Experiments: Virtual Network Simple topology

All VMs connected to the same bridge

ClientGuest 1

bridge

ClientGuest 2

ClientGuest 3

ClientGuest 4

ClientGuest 5

ServerGuest 1

ServerGuest 2

ServerGuest 3

ServerGuest 4

ServerGuest 5

Experiments: Virtual Network Simple topology

All VMs connected to the same bridge

Up to 16 virtual links IPerf TCP channels Example with 7 links

ClientGuest 1

bridge

ClientGuest 2

ClientGuest 3

ClientGuest 4

ClientGuest 5

ServerGuest 1

ServerGuest 2

ServerGuest 3

ServerGuest 4

ServerGuest 5

Experiments: tests SMP disabled SMP enabled Static domain scheduling

10 iterations for each experiment 1 minute per link Samples every 5 sec Average value

Experiments: Results NoSMP vs. SMP

Experiments: Results Dynamic scheduling vs Static scheduling

Model

Model: assumptions Simple resource model

Single type of resource Resources completely separated in system and

network Network described by the number of virtual links Bandwidth equally distributed among links

Model M: maximal total bandwidth M – K: minimal total bandwidth (n): total bandwidth

Bandwidth

Network resources

System resources

M

KTotal

resources

n links

Model Model curve vs. experimental data: error less than

2%

Security mechanisms

Security mechanisms Adding security brings

More workload More networking

We focused on increase of number of links (eg. firewalls)

Security mechanisms Number of links increases by a factor s

Depending on topology Depending on the security mechanism

The model allows prediction on the loss of bandwidth

Model application 1/2 Scenario: server consolidation

Computation power available The virtual network must supply the physical interface If the virtual network is well-designed, the virtual

network supports the transaction

Model application 2/2 What happens if we introduce a firewall? Applying the model we can esteem the resulting

bandwidth

Conclusions

Future works Improve the model

Relax assumptions Forecast parameters without experiments

Validate the model Other architecture Other security solutions

Improve Xen D2D communication Optimization

Conclusions We developed a simple (but still effective) model

Explain how virtual network works in Xen Foresee performance of the virtual network

Planning Impact of security solutions

We show the limits of current Xen’s implementation and suggested improvements

Thank you

Any question?