The document discusses performance tuning in Xen virtualization, detailing its architecture, virtualization modes, and the importance of selecting the right mode based on workload. It emphasizes the role of the control domain (dom0) and various performance tuning options for domains, including memory management and CPU pinning. The conclusion highlights the necessity of experimenting with different modes for optimal performance and staying updated on enhancements.

1. Performance tuning Xen
Roger Pau Monn´
e
roger.pau@citrix.com
Antwerp – 8th of April, 2013

2. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Xen Architecture
Control Domain
NetBSD or Linux
device model
(qemu)
toolstack Paravirtualized
(PV) Fully
Domain: Virtualized
NetBSD or Linux (HVM)
Domain:
Hardware netback netfront Windows,
Drivers blkback blkfront FreeBSD...
Xen Hypervisor
I/O Devices CPU Memory Hardware
Antwerp – 8th of April, 2013 Performance tuning Xen 2 / 27

3. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Paravirtualization
Virtualization technique developed in the late 90s
Designed by:
XenoServer research project at Cambridge University
Intel
Microsoft labs
x86 instructions behave differently in kernel or user mode,
options for virtualization were full software emulation or
binary translation.
Design a new interface for virtualization
Allow guests to collaborate in virtualization
Provide new interfaces for virtualized guests that allow to
reduce the overhead of virtualization
The result of this work is what we know today as
paravirtualiztion
Antwerp – 8th of April, 2013 Performance tuning Xen 3 / 27

4. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Paravirtualization
All this changes lead to the following interfaces being
paravirtualized:
Disk and network interfaces
Interrupts and timers
Boot directly in the mode the kernel wishes to run (32 or
64bits)
Page tables
Privileged instructions
Antwerp – 8th of April, 2013 Performance tuning Xen 4 / 27

5. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Full virtualization
With the introduction of hardware virtualization extensions
Xen is able to run unmodified guests
This requires emulated devices, which are handled by Qemu
Makes use of nested page tables when available.
Allows to use PV interfaces if guest has support for them.
Antwerp – 8th of April, 2013 Performance tuning Xen 5 / 27

6. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
The full virtualization spectrum
VS Software virtualization Poor performance
VH Hardware virtualization Room for improvement
PV Paravirtualized Optimal performance
es ns
ge str rd
bl tio
s
a
er
bo
uc
tim
k
an ge her
or
tw
ot
d
ta
pa in
an
m
ne
d d
ts
ed
d
up
ile
an
at
rr
iv
ul
isk
te
Pr
Em
In
HVM D
VS VS VS VH
HVM with PV drivers PV VS VS VH
PVHVM PV PV VS VH
PV PV PV PV PV
Antwerp – 8th of April, 2013 Performance tuning Xen 6 / 27

7. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Guest support
List of OSes and virtualization support:
PV PVHVM HVM with PV drivers HVM
Linux (PVOPS) YES YES YES YES
Windows NO NO YES YES
NetBSD YES NO NO YES
FreeBSD NO NO YES YES
OpenBSD NO NO NO YES
DragonflyBSD NO NO NO YES
Antwerp – 8th of April, 2013 Performance tuning Xen 7 / 27

8. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Kernbench
Kernbench
Results: percentage of native, the lower the better
140
135
130
125
120
115
110
105
100
95
90
PV on HVM 32 bit HVM 32 bit PV 32 bit
PV on HVM 64 bit HVM 64 bit PV 64 bit
Antwerp – 8th of April, 2013 Performance tuning Xen 8 / 27

9. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Specjbb2005
SPECjbb2005
Results: percentage of native, the higher the better
100
90
80
70
60
50
40
30
20
10
0
PV 64 bit PV on HVM 64 bit
Antwerp – 8th of April, 2013 Performance tuning Xen 9 / 27

10. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Iperf
Iperf tcp
Results: gbit/sec, the higher the better
8
7
6
5
4
3
2
1
0
PV 64 bit PV on HVM 64 bit PV on HVM 32 bit PV 32 bit HVM 64 bit HVM 32 bit
Antwerp – 8th of April, 2013 Performance tuning Xen 10 / 27

11. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
What virtualization mode should I choose?
Linux supports several virtualization modes, which one is
better?
Depends on the workload.
Generally PV mode will provide better performance for IO,
but when using 64bit guests PV can be slower.
There isn’t a fixed rule here, the best way to find out is to
evaluate the workload on the different kind of guests.
Antwerp – 8th of April, 2013 Performance tuning Xen 11 / 27

12. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Dom0
Dom0 is the most important guest in the Xen infraestructure.
It can become a bottleneck easily if not configured correctly.
Dom0 is in charge of creating the guests, but usually also
provides the backends and device models for guests.
Xen provides some options to tune performance of Dom0
Antwerp – 8th of April, 2013 Performance tuning Xen 12 / 27

13. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
dom0 mem boot option
dom0 mem tell Xen how much memory can be used by the
Dom0.
If not set all memory will be assigned to the Dom0, and
ballooning will be used when launching new guests, reducing
the memory used by the Dom0.
The value should be set depending on the usage, HVM guests
consume more memory in the Dom0 because they need a
Qemu instance.
If dom0 mem is set make sure to disable ballooning in the
toolstack by setting autoballoon=0 in /etc/xen/xl.conf.
Antwerp – 8th of April, 2013 Performance tuning Xen 13 / 27

14. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
dom0 max vcpus and dom0 vcpus pin
dom0 max vcpus: maximum number of CPUs the Dom0 will
see, also depends on the utilization of the Dom0 and the type
of guests.
dom0 vcpus pin: pinning Dom0 vcpus to physical CPUs is a
good idea for systems running IO intensive guests.
Setting up the serial cable: although not important for
performance, setting up a serial cable is really important when
debugging. For more info:
http://wiki.xen.org/wiki/Xen_Serial_Console
Antwerp – 8th of April, 2013 Performance tuning Xen 14 / 27

15. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Dom0 Boot tunning options
For example if I had to set up a Dom0 on a machine with 8
CPUs and 8GB of RAM I would use the following boot line:
com1=115200,8n1 console=com1 dom0 mem=1024M
dom0 max vcpus=2 dom0 vcpus pin.
More info about boot parameters can be found at:
http://xenbits.xen.org/docs/unstable/misc/
xen-command-line.html.
Antwerp – 8th of April, 2013 Performance tuning Xen 15 / 27

16. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
General performance notes
Always try to use physical disks as backends. Xen has mainly
two ways of connecting disks to the guest depending on the
format of the image, if it’s a block device it will be attached
using blkback, which is inside the Linux kernel and it’s faster.
Take into account the number of CPUs your physical box has
and avoid using more VCPUS than PCPUS if running
performance intensive applications.
Antwerp – 8th of April, 2013 Performance tuning Xen 16 / 27

17. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Pinning CPUs
You can pin VCPUs to PCPUs in order to obtain better
performance or to distribute the workload across your CPUs
to suit your needs. For example low latency VMs can be
exclusively pinned to different PCPUs.
cpus: allows to select in which CPUs the guest can run. The
list can also contain specific CPUs where the guest is not
allowed to run. Specifying ”0-3,5,ˆ1” allows the guest to run
on CPUs 0,2,3,5.
If Dom0 is pinned to certain PCPUs avoid running guests on
those PCPUs to obtain better performance. If Dom0 is pinned
to CPU 0, use the following CPU mask in order to prevent
other guests from running on CPU 0: ”ˆ0”.
Antwerp – 8th of April, 2013 Performance tuning Xen 17 / 27

18. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Scheduler options
The Xen scheduler has several options that can also be
tunned from the guest configuration file, in order to give a
certain guest more share from the processor or to schedule it
more frequently.
cpu weight: weight of the domain in terms of CPU utilization.
For example a domain with a weight of 512 will get twice as
much CPU than a domain with a weight of 256. Values range
from 1 to 65535.
cap: fixes the maximum amount of CPU a domain is able to
consume. Expressed in percentage of one physical CPU. 100
is one CPU, 50 half a CPU, 400 four CPUs.
More info can be found at http:
//xenbits.xen.org/docs/unstable/man/xl.cfg.5.html
Antwerp – 8th of April, 2013 Performance tuning Xen 18 / 27

19. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Driver Domains I
Control Domain
NetBSD or Linux
device model
(qemu)
toolstack Paravirtualized
(PV)
Domain:
NetBSD or Linux
Hardware netback netfront
Drivers blkback blkfront
Xen Hypervisor
I/O Devices CPU Memory Hardware
Antwerp – 8th of April, 2013 Performance tuning Xen 19 / 27

20. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Driver Domains II
Control Domain
NetBSD or Linux
device model
(qemu)
toolstack Paravirtualized
Driver Domain
(PV)
Domain:
NetBSD or Linux
Hardware netback netback netfront
Drivers blkback blkback blkfront
Xen Hypervisor
I/O Devices CPU Memory Hardware
Antwerp – 8th of April, 2013 Performance tuning Xen 20 / 27

21. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Driver Domains III
Driver domains allow to offload work normally done in Dom0
to other domains.
It also provides better security, less surface for exploits in
Dom0.
This is a current work-in-process.
Antwerp – 8th of April, 2013 Performance tuning Xen 21 / 27

22. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
HVM specific I
HVM guest require the usage of assisted paging, in order for
the guest to see the memory area as contiguous when it’s not.
HAP: (Hardware Assisted Paging) is used by default since it
tends to perform better under most workloads
shadow: was introduced before HAP, and can provide better
performance under certain workloads that have low TLB
locality (for example databases or java applications).
Again, the best way to know is to try the workload by yourself.
Antwerp – 8th of April, 2013 Performance tuning Xen 22 / 27

23. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
HVM specific II
HVM domains require a Qemu instance in Dom0 to perform
the necessary device emulation.
This might be a bottleneck if running a lot of HVM domains
in the same node, since each one requires a Qemu instance
running in Dom0 that uses both Dom0 CPU and Memory.
To avoid this, we can launch the Qemu process in a different
domain called ”Stubdomain”.
This allows to offload work from Dom0.
Antwerp – 8th of April, 2013 Performance tuning Xen 23 / 27

24. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
HVM specific III
Control Domain
NetBSD or Linux
device model
(qemu)
Fully
toolstack Virtualized
(HVM)
Domain:
Windows,
FreeBSD...
Hardware netback
Drivers blkback
Xen Hypervisor
I/O Devices CPU Memory Hardware
Antwerp – 8th of April, 2013 Performance tuning Xen 24 / 27

25. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
HVM specific IV
Control Domain
NetBSD or Linux
device model
(qemu)
Fully
toolstack stubdomain Virtualized
(HVM)
MiniOS Domain:
Qemu Windows,
FreeBSD...
Hardware netback
Drivers blkback
Xen Hypervisor
I/O Devices CPU Memory Hardware
Antwerp – 8th of April, 2013 Performance tuning Xen 25 / 27

26. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Conclusions
Xen offers a wide variety of virtualization modes.
The best way to know which mode will bring better
performance is to try it, although there are several tips that
apply to all guests.
We are constantly working on performance improvements, so
keep updated in order to get the best performance.
Antwerp – 8th of April, 2013 Performance tuning Xen 26 / 27

27. Xen Architecture Xen virtualization modes Support in OSes Dom0 tunning Specific VM options Conclusions
Q&A
Thanks
Questions?
http://wiki.xen.org/wiki/Xen_Best_Practices
http://wiki.xen.org/wiki/Xen_Common_Problems
Antwerp – 8th of April, 2013 Performance tuning Xen 27 / 27