1

CPU Allocation on Xen Virtual Networks

Rodrigo S. Couto, Hugo E. T. Carvalho, Lyno Henrique G. Ferraz,
Miguel Elias M. Campista, Luı́s Henrique M. K. Costa and Otto Carlos M. B. Duarte
Grupo de Teleinformática e Automação - PEE/COPPE - DEL/POLI
Universidade Federal do Rio de Janeiro - Rio de Janeiro, Brazil
Email:{souza,hugo,lyno,miguel,luish,otto}@gta.ufrj.br
Abstract—In this work we investigate virtual network per- (TRs) passing through a traffic forwarder (TF). All machines
formance using Xen. We perform experiments using scheduling run Debian Linux and TF runs Xen 3.4.2 system.
parameters and static CPU allocation. Results indicate that the
forwarding packet rate of each virtual router can be controlled A. Domain 0 CPU cores
by adjusting CPU scheduling parameters. All packets destined to virtual routers must pass through
I. I NTRODUCTION Domain 0. Thus, we need first to evaluate the performance
The growing need to develop pluralist architectures for the of Domain 0 adjusting the number of CPU cores allocated
future Internet leads to the use of virtual routers in platforms to it. In this experiment, we send packets from one of the
such as Xen [1]. Xen implements a software layer called VMM TGs to one of the TRs at a fixed rate and vary the number
(Virtual Machine Monitor) which controls how virtual ma- of CPU cores allocated to Domain 0. We run this test for
chines, called Domains, access hardware resources. Domain 0 up to four virtual machines forwarding packets and sharing
is a special Domain that implements an interface between the same TF physical network interfaces. On the one hand,
abstract devices within virtual machines and physical devices. the performance of packet forwarding is severely impacted
In the case of networking, packets destined to unprivileged when all Domains share one CPU core because of the high
Domains must pass through Domain 0 before reaching virtual contention for CPU resources. On the other hand, the best
interfaces. This communication between Domain 0 and other performance is achieved when Domain 0 has one exclusive
Domain imposes limitations to network tasks due to high CPU core. Surprisingly, assigning more cores to Domain 0 results
consumption. Moreover, in Xen environment the Domains in performance degradation in our scenario. This result shows
share physical resources, including CPU. In this paper, we that the tasks that Domain 0 executes to forward packets
aim at analyzing the impact of CPU resource allocation on through virtual machines are single-threaded and that these
virtual routers performance. tasks are not well suitable for multi-core environments. The
behavior observed in this test is independent of the number of
II. X EN C REDIT S CHEDULER
virtual machines concurrently forwarding packets.
For each Domain running on a Xen environment is assigned
B. Cap Adjustment
a number of Virtual CPUs (VCPUs). These VCPUs correspond
Credit Scheduler parameters can be used to offer different
to the CPUs that the running processes in a Domain can see.
service levels to virtual routers. In our case, we use the cap
Xen implements the Credit Scheduler to manage CPU time
to show how packet rate can be adjusted when the system
given to each Domain. This scheduler dynamically allocates
administrator has control of CPU resource sharing among
a physical CPU core to each VCPU for a certain amount of
virtual routers. The experiment consists in sending packets
time. In Xen platform the administrator can also manually
from a TG to TR at a fixed rate through each virtual router
allocate a CPU core to a VCPU of a certain Domain but cannot
in TF. In this experiment, TF has two virtual routers with
guarantee if another VCPUs will share the same core. The
each one having its own network interface pair and sharing
basic idea of Xen Credit Scheduler is to assign credits to each
the same CPU core. Thus, we ensure that the main resource
Domain based on two parameters: weight and cap. Each time a
that the virtual routers are sharing is the CPU. Domain 0 in
Domain executes, i.e. each of its VCPUs, it consumes credits.
this case has its own CPU core.
Based on credit accounting, the scheduler determines which
In this experiment we assign a fixed cap of 100 to a
VCPU can run. The weight determines the share of CPU time
virtual machine (VM2) and vary the cap of the other virtual
each Domain can get and doesn’t represent an upper-limit, i.e
machine (VM1). Results show that the performance of VM1
Domains can use more than its share when CPU is idle. The
decreases when its cap is reduced whereas VM2 performance
cap, in turn, is an absolute value that represents a percentage
increases. Thus, we show that we can assure certain perfor-
of CPU a Domain can use, i.e cap=50 is half CPU. The cap
mance level to each virtual router, merely adjusting the cap.
gives more control to the system administrator about the CPU
Our future work aims at building a mechanism to dynamically
usage by the Domains.
adjust caps depending on each Domain performance level.
III. P ERFORMANCE E VALUATION
R EFERENCES
We aim at analyzing how Domain 0 CPU allocation impacts
[1] P. Barham, B. Dragovic, K. Fraser, S. Hand, , T. Harris, A. Ho, R. Neuge-
forwarding performance and how can a system administrator bauer, I. Pratt, and A. Warfield, “Xen and the art of virtualization,” in
manage CPU resources to guarantee a certain throughput to ACM SOSP, Oct. 2003, pp. 164–177.
a virtual router. To accomplish this, we conduct two experi- ACKNOWLEDGMENT
ments. These experiments were carried out in a testbed with This work was supported by CNPq, CAPES, FINEP, FUNTTEL, and
traffic generators (TGs) sending packets to traffic receivers FAPERJ.