Patagonia -
       A Dual Use Cluster of PCs for Computation and Education

    Felix Rauch, Christian Kurmann, Blanca Maria Müller-Lagunez
                         and Thomas Stricker
                 Laboratory for Computer Systems ETH
                    Swiss Institute of Technology
                     CH-8092 Zürich, Switzerland
                {rauch,kurmann,bmc,tomstr}@inf.ethz.ch 



Computer clusters in major research universities are used for research
and education. While both kinds of installations can be called
``clusters'' in English, the installations are far from looking
identical.  The PCs of an education cluster are workstations that are
sparsely distributed across tables in a classroom and do have screens,
keyboards and mice, while clusters of PCs for high performance
computing are densely packed rack mounted systems kept in a cooled
machine room and wired to one central operator console. Despite the
completely different look of the two kinds of clusters, the current
trends in technology mandate that they are built with nearly the same
components, namely commodity PCs leveraging off an ever increasing
compute performance found in single chip microprocessors, high volume
DRAM memories and low cost disk drives of several GB capacity.
Besides the technical specifications, both types of PC clusters also
have in common that they have a low utilization compared to
traditional mainframes and supercomputers. Typically the education
clusters are only used during the day and maybe in the evenings
whereas the compute clusters are often idle while the programmers work
hard to improve the code of their parallel applications. We admit that
in both usage models the management could fill the capacity of the
idling processors with embarrassingly parallel computations in
cryptography, number theory or combinatorics for research clusters or
alternatively bring in PC users from the street that are looking for
free text processors or internet access - but such complementary use
would hardly improve the cost effectiveness of the installations.

As a consequence we initiated the Patagonia cluster project at ETH
Zürich to build a cluster that fits both the needs of education and
the needs of research. While education has priority during the day on
our cluster, research has priority during the night and with some
limitations during vacations. The Patagonia cluster is a collection of
high-end PCs in a classroom interconnected by a fully switched Gigabit
Ethernet. Dual processors, large memories and fast disks make the
machines fit for experiments in parallel and distributed computing
while fast graphics cards, 17 inch monitors, mice and keyboards make
them suitable for instruction in our advanced CS curriculum.

While our ideas and the corresponding experimental study clearly
originates from a university environment we would like to mention the
striking similarities to many modern corporate computing environments
using PCs. In those environments the computing needs are quite
similar. Most companies rely on a rapidly growing number of compute
intensive tasks in data mining, combinatorial optimization and process
simulations in addition to the typical personal computing needs of a
large number of employees at their desks. A usage model that can cover
personal computing and occasional compute-intensive tasks with a
single infrastructure would be a great advantage to those
businesses. However such a usage model explicitly excludes the
traditional transaction processing because it has extremely high
requirements regarding reliability and data security that clusters of
PCs can not meet at this time.

In our subsequent contribution to the workshop we will describe the
requirements and how we managed to find a common system software basis
for the dual use cluster based on a multi-boot setup of several
carefully isolated operating system images. We were successful in
installing a system software setup with several fully sandboxed
operating systems that are bootable in a networked environment. The
setup includes multiple Windows NT installations which contradicts the
recommendations of the manufacturer, who is claiming technical
unfeasibility. In our Patagonia project we particularly focus on the
benefits of the high end facilities provided by the research part of
the cluster (mainly the high speed disks and the high speed network)
that also help tremendously with the installation and maintenance of
the education partitions of the cluster. Although the true cost of
such an installation based on university pricing remains very hard to
estimate, we are convinced that the combined installation could be
cheaper and easier to maintain than the installation of two
specialized clusters for scientific computing and for education.