Capacity of desktop clouds for running HPC applications: A revisited analysis

Abstract

Desktop Clouds, such as UnaCloud and CernVM, run scientific applications on desktop computers installed in computer laboratories and offices. These applications run on virtual machines using the idle capacities of that desktops and networks. While some universities use desktop clouds to run bag of tasks (BoT), we have used these platforms to run High Performance Computing (HPC) applications that require coordination among the nodes and are sensible to communication delays. There, although a virtual machine with 4 virtual cores on computers released in 2012 may achieve more than 40 GFLOPs, the capacity of clusters with tens or hundreds of virtual machines cannot be determined by multiplying this value. In a previous work, we studied the capacity of desktop clouds for running applications non-intensive on communications on our computer labs. This paper presents a revisited analysis, focused on the capacity of desktop clouds for running HPC applications. The resulting information can be used for researchers deciding on investing on HPC clusters or using existing computer labs for running their applications, and those interested on designing desktop clusters that may achieve the maximum possible capacity.