Speaker
Description
An innovative, distributed, and elastic computing ecosystem, called UniNuvola,
is being deployed at the University of Perugia, involving the Department of
Chemistry, Biology and Biotechnologies, the Department of Physics and Geol-
ogy, the Department of Mathematics and Informatics and the Department of
Engineering. The aim of the project is the creation of a federated and scalable
computing infrastructure, providing scientific services to end users belonging
to both academic structures and, in perspective, SMEs. It represents a proof
of concept of a distributed computing infrastructure empowered with software-
defined networking capabilities, with every node laying transparently its virtual
resources, services, and microservices on a virtual backbone. Our main objective
is to design a virtual distributed networking infrastructure to federate an inno-
vative architecture, with dynamical resource allocation, intelligent management
of large volumes of data and compliant with present European federated com-
putation paradigm and data protection policies. The federation also manages
a heterogeneous collection of virtual application environments by organizing
them into well tested and auto-consistent packages, ready to be used by orga-
nizations guaranteeing, at the same time, the requested performance and result accuracy features. To this end, a first prototype of four Dell Power Edge R940
nVME servers, each equipped with 2 Intel Xeon Gold CPU, 512 GB of RAM,
and 16 TB of disk space, has been configured with the most recent software
solutions for pursuing the above mentioned objectives. More in detail, a Kuber-
netes cluster has been installed, using the Rook operator for deploying a CEPH
scalable distributed storage. We have also investigated the adoption of both
Metallb and OVN load balancers. User authentication has been managed with
the Vault server interfaced to the University LDAP, while Jupyter Hub has been
containerized into Kubernetes to serve Notebooks for multiple users. On the
other hand, for those workloads already running in virtual environments that
are difficult to containerize into Kubernetes pods, KubeVirt technology provides
us with the possibility of enabling KVM-based virtual machine workloads to be
managed as pods.
Upon this (virtual) infrastructure a ready-to-use collection of scientific pack-
ages, for both research and education purposes, is being developed. To this
end, the capabilities of UniNuvola will be firstly benchmarked with various use
cases, built upon computational chemistry and machine learning applications.
Computational chemistry applications are widely recognized for their high de-
mands on CPUs and storage, making them the ideal candidates for testing the
scalability of the architecture. The significance of machine learning lies not
only in its wide range of applications but also in its high memory requirements.
Applications ranging from image recognition to intrusion detection algorithms,
tested on well-known datasets or from live interaction with external sources, will
allow benchmarking the capability of the platform with respect to the commercial,
widespread alternatives. In addition, both cases can be utilized to test future
improvements to the infrastructure, such as the inclusion of GPUs and quantum
computing. In fact, during the second phase of the project, both high-end nodes
with GPUs (Nvidia A100) and a solid-state SpinQ Triangulum NMR quantum
computer will be integrated into the UniNuvola cluster to realize an academic
across-the-board data center able to serve a variety of instances.
