MUGASA HATWIB
The Glasgow Raspberry Pi Cloud:
A Scale Model for Cloud Computing Infrastructures
Cloud Data Center Development/Testbed Environments
A typical Cloud data center has tens and thousands of servers and making it very
expensive for educational and research purposes. Most research in Data Centers using
Cloud computing is by uses testbed/development environments with a handful of
machines. The typical development environment that consists of a couple of servers is
still very expensive for most research and educational institutions considering the data
center space, power, cooling, etc.
Cloud Data Center Simulators
Data center Cloud simulators have been developed to provide a low cost alternative to
the development environments. In the past, simulators have been used successfully to
model state-of-a-target systems however there are essential Cloud Computing
properties that these simulators have failed to capture. Traffic patterns in operational
Cloud data center networks constantly change over time and are generally
unpredictable in the long term, something that cannot be fully modelled in a simulated
environment.
Simulation also does not model cross layer correlation between application, network
and virtualization. An example is the iCloud simulator of the Cloud infrastructure aimed
at simulating instance types provided by Amazon without considering the underlying
network behavior.
Furthermore, with Cloud simulation tools like iCanCloud, CloudSim and GreenCloud still
in there infant development and physical Cloud Dev environments running on x86
processors being very expensive for low-budget research project, a novel alternative is
to use a cutting-edge “scale-model” of a Clould system to create a highly reliability and
affordable Dev environment.
PiCloud as an alternative scaled Cloud DC model
The low-power, low-cost Raspberry Pi provides an affordable option to construct a
miniature Cloud Data Center. It also allows for the reproduction of actual traffic patterns
with realistic cloud applications. The applications and technics developed suggested in
this research are readily adapted to production Cloud environments
This research provides an insight into the Glasgow PiCloud project which uses
Raspberry Pi devices in a cluster to build a scaled model of a Cloud data center. It
covers the areas of system design, research direction, discussions on the Raspberry Pi
device and related works.
The System Design
The prototype scale model of the PiCloud cluster was built using 56 Model B version
Raspberry Pi devices housed in a rack constructed from Lego bricks. In order to reflect
a typical data center network architecture, the Raspberry Pi devices are interconnected
through a canonical multi-root tree topology. Machines in the same rack are connected
to the same Top Of rack switch, which in turn connects to the rest of the topology
through an OpenFlow enabled aggregation switch thus allowing it to be fully
programmable and compatible with the leading-edge Software Define Networking
(SDN) research.
The accessibility and design of the prototype means the PiCloud clusters can easily be
re-cabled to form a fat-tree topology. In a fat-tree topology, the links become "fatter" as
traffic moves up the tree towards the root. By choosing the fatness of the links, the
network can be tailored to efficiently use any bandwidth made available by packaging
and communications technology.
Figure 1 below show the physical Raspberry Pi devices that were used in this research
and Figure 2 shows the System Architectural design.
Figure 1.
Figure 2
System Architecture
The software stack for an individual Pi is shown in Figure below
The Raspberry Pi devices used in this research where run using Linux on a 16GB
Sandisk SD card storage.
The Linux Operating System that was used is the Raspbian, this is a Debian flavor of
Linux that has been optimized to run on the Raspberry Pi hardware through
contributions from the Raspberry Pi community. Raspbian OS comes with over 35,000
pre-compiled software packages built with support for the Raspberry Pi hardware
floating point capacities.
It is also installed with the Linux Container Suite of programs to provide operation
system level virtualization, playing a role similar to the hypervisor in x86 virtualization
technologies. Currently, they were able to comfortably support three containers
concurrently on a single Raspberry Pi. There is also an API daemon on each Pi to
provide a RESTful management interface for facilitating virtual host management and
interacting with a head node or the pimaster.
System Virtualization
The most popular virtualizations tools on the market like Xen and VMWare are very
memory intensive to be run on a 256MB RAM device, thus rendering them useless for
the PiCloud project. In addition, most of the architectural support for virtualization to
efficiently isolate guest operating systems from the host is provided for x86 CPU
processors. There is an ongoing effort by Xen to provide ARM platform support however
this was still under development when the PiCloud project was being run.
As an alternative, Linux Containers were used for virtualization, however, they did not
provide the level of isolation that full virtualization achieves. They only provide a virtual
environment that has its own process and network space.
Management API
As part of the LXC suite of packages, there is a set of tools that can be used as a
management layer to administer the PiCloud, however, they are not fully functional on
the Pi platform. Therefore a webserver on the pimaster was used to provide a webbased control panel for users and administrators to manage the PiCloud as shown in
Figure 4 below. Through the interaction with daemons running on individual Pi devices
and the pimaster, the website is used to control workloads using RESTful interfaces.
Figure 4
Research directions
The PiCloud can be used for experimenting with new algorithms for Virtual Machine
(VM) management and directly observing resulting behaviors on all layers of the Cloud
architecture, something that would not be achieved with over-simplified assumptions of
simulators. By operating an actual physical infrastructure, the improvement of file
management and migration techniques can be adequately evaluated. Observations on
the interaction and effects of specific optimizations on the cloud configurations can be
made at all different layers.
The PiCloud project can also be used to investigate ways of reducing network
congestion through improved resource allocation, as well as looking at novel network
architectures and technologies that require significant changes to the infrastructure.
Through using physical devices to build the Cloud DC, crucial administration
management requirements like API design, UI design etc., will not be overlooked or
underestimated compared to if the Cloud DC was built using simulators.
A future development in the Cloud computing is the adjustment or removal of
virtualization techniques. Like with the PiCloud project, this could be achieved by
switching the virtualization tools to lighter mechanisms like the Linux Containers.
Alternatively, this can be done by removing virtualization completely and renting out
physical nodes.
Conclusion
The PiCloud project showed that a Cloud DC can be hosted without the extensive
physical footprint in terms of expensive power and cooling management, which cover
33% of total power consumption, by scaling down the hardware capacity. However this
is not the same for software capacity and limiting the services that can be applied on the
PiCloud to a few lightweight applications like http server, Hadoop, etc.
The table below showed the cost breakdown of a development environment compared
PiCloud
Server Cost Power
Cooling
Development
Environment
$112,000 10,080W/h Yes
PiCloud
$1,960 196W/h
No
With the reduction in hardware prices and increase in capacity, and if the trend
continues, there is anticipation that Cloud computing will be more affordable for
research and education deployments.
References:
[1] Fung Po Tso, David R. White, Simon Jouet, Jeremy Singer, Dimitrios P. Pezaros,
"The Glasgow Raspberry Pi Cloud: A Scale Model for Cloud Computing Infrastructures,"
icdcsw, pp.108-112, 2013 IEEE 33rd International Conference on Distributed
Computing Systems Workshops (ICDCSW), 2013
[2] A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel, “The cost of a cloud: Research
problems in data center networks,” ACM SIGCOMM Computer Communication Review,
vol. 39, no. 1, January 2009.
[3] “Raspbian.” [Online]. Available: http://www.raspbian.org/
[4] “Raspberry pi.” [Online]. Available: http://www.raspberrypi.org