ASP projects

SSP: Switched Positive Dynamical Systems for Cloud Computing Load Balancing

The project focuses on a very important challenge currently arising in Cloud infrastructure design, namely the problem of optimal allocation of load among virtual machines. Efficient load balancing in a Cloud environment can indeed lead to important benefits for both end users and service providers. In particular, end users can experience a better quality of service in terms of shorter response time and higher availability, whereas service providers can more efficiently exploit their software/hardware resources and reduce costs by energy saving. This latter observation perfectly fits the characteristics of so-called green ICT. The scenario considered in the project consists of a cluster of servers which have to perform a set of jobs according to a timevarying input flow of requests. The number of active servers is variable since they might be switched on in order to cope with sudden peaks of requests and turned off to reduce energy consumption during idle periods. In addition, jobs should be uniformly allocated among active servers so as to provide a fair and balanced use of resources from the users’ viewpoint. The queue dynamics of service requests lends itself to be described by differential equations involving positive variables.
Moreover, the changes in system configuration call for a description in terms of switched dynamical systems. The performance index to be optimized is a trade-off between the number of active servers (which should be kept small to reduce energy consumption) and response time (which might lead to penalties in case of violation of Service Level Agreements). The project considered a simplified, yet realistic, mathematical model of the overall system. This model is described by a feedback solution which provides, at each time point, the number of active servers and their load share. The proposed solution was simulated in MATLAB®, implemented as a Software as a Service on the Microsoft Azure Cloud infrastructure, and tested through a set of benchmarks. The numerical and experimental results showed a good level of agreement. Moreover, measurements resulting from the proposed algorithm were compared with other (classical) resource allocation strategies. As a result, it is possible to affirm that the model-based approach offers a promising alternative to the present state of the art. Besides the technical aspects, the team worked hard on tackling all the different aspects of the problem, including market analysis and research on current trends in Cloud computing technologies. The various educational backgrounds of members, integrating modeling, simulation, control and computer science, contributed to the success of the project.

Principal Academic Tutor
Patrizio Colaneri
Electronics, Information and Bioengineering, Politecnico di Milano

Academic Tutors
Paolo Bolzern
Electronics, Information and Bioengineering, Politecnico di Milano
Marco Gribaudo
Electronics, Information and Bioengineering, Politecnico di Milano

External institution
CNR, Istituto Nazionale di Fisica
Nucleare

Team members
TEAM A
Marco Agnese [Team Controller], Mathematical Engineering
Riccardo Cipolleschi [Project Communication Coordinator], Computer Engineering
Luigi Colangelo, Aerospace Engineering
Daniele Cozzi, Mathematical Engineering

Download the poster of the Project