Building an IaaS cloud with droplets: a collaborative experience with OpenStack

Abstract

The goal of a collaborative cloud is to promote resource sharing among different institutions using existing computational infrastructure. To accomplish that, we propose the PID (Platform for IaaS Distribution) cloud, where users from research institutions can benefit from IaaS services with minimum management and maintenance effort and with no subscription fees. The only requirement is to share at least one server capable of running virtual machines. To build PID, we rely on the OpenStack orchestrator as the basis of our implementation and, consequently, modifications are needed to accommodate the geo-distributed architecture. We present such modifications, which are concentrated on a different virtual machine scheduling, new hierarchical roles for users, and adapted web interfaces for management and control. Based on these modifications, we deploy a prototype in three universities in Rio de Janeiro. We experimentally evaluate PID regarding the latency between sites and the scalability according to the number of servers and virtual machines. Based on these experiments, we can estimate that an OpenStack cloud with a few hundred servers with tens of virtual machines each would add, on average, a few Mb/s of control traffic toward the Controller. We then conclude that the amount of control traffic generated by OpenStack services is not enough to surpass network capacities of our academic sites. Our results also show that the additional latency expected in a geo-distributed cloud has little or no impact on the OpenStack network component, called Neutron.