Information Technology Reference
In-Depth Information
The first group is concerned with measuring the performance degradation (over-
head) of using virtual machines over desktops for the execution of applications in
comparison with the direct use of the physical environment, so called native. In [15]
the authors compare the floating-point performance of VMware Player virtual ma-
chines by using Windows and Linux as host and guest operating systems on the AMD
Dual-Core architecture. They conclude that both operating systems used as guests
induced performance penalties compared to a native execution, but as a guest operat-
ing system Linux delivers better performance than Windows. Since VMware Player
emulates only the kernel-mode instructions, the authors hypothesize that incidence of
such instructions in Linux is lower than in Windows.
In [7] the authors compare the performance of executions on native versus virtua-
lized environments by using benchmarks based on CPU intensive tasks. They found
that overhead induced by virtualization is less than 10%. The experiments are set on
architectures based on the Intel Dual Pentium III processor by using VMware
Workstation as hypervisor and RedHat as host and guest operating system.
A second group, which is more related to our research, is concerned with eva-
luating the intrusiveness of virtual machines executed as low-priority processes on a
desktop in relation to the performance perceived by an end-user that is simulta-
neously using its environment. In [5] the authors evaluate the performance of vir-
tual environments based on the hypervisors VMware Player, QEMU, VirtualPC,
and VirtualBox on Intel Dual-Core processors by using Windows as host and Linux
as guest operating systems. The results show marginal performance impact in the
presence of a single virtual machine as long as only single-threaded applications run
at the host operating system. In contrast, multi-threaded applications running at the
host operating system suffer a considerable performance drop which ranges from
10% to 35% compared with the same execution on the native environment. On the
other hand, for the applications executed on the virtual environment they found that
performance depends on the application type and the virtualization software used.
Indeed, for CPU intensive tasks the overhead revolves around 15% to 30% which is
considered acceptable. However, disk IO and network IO performances are severely
penalized, and thus the authors suggest not using virtual environments for such
execution scenarios.
In [3] the authors evaluate the intrusiveness of a Desktop Grid system named Un-
aGrid. For this purpose, they execute the hypervisor type 2 VMware Workstation over
an Intel Core 2 Duo processor. They simultaneously execute CPU intensive tasks on
two operating systems, Linux as guest and Windows as host. Each virtual machine is
executed as a low-priority process in background and it is assigned with one and two
cores subsequently. In such tests, the results show performance degradation of less
than 1%. The authors conclude that priority set to the virtual environment at the oper-
ating system level allows an exclusive harvesting of idle CPU resources that guaran-
tees very low impact on the performance perceived by end-users.
This research complements and extends the work presented above by benchmark-
ing a Desktop Grid based on virtualization over several generations of modern
processor architectures in order to evaluate variations in the intrusiveness caused to
