Data Driven Server Allocation at Virtual Computing Labs

Abstract

Virtual computing labs (VCL) are cloud computing platforms that provide users remote access to software applications. In this paper, we develop a data driven approach for server allocation in a VCL. The main challenge is to decide how many servers should be preloaded with which applications, and how many servers should be left flexible, to be loaded with the requested applications on demand. If a preloaded server with a desired application is available, the user gets immediate access. If not, the user gets delayed access after some extra loading time, if a flexible server is available. If no server (dedicated or flexible) is available, the user is blocked. We measure the service quality by the fractions of users who get immediate or delayed access, and the system cost by the number of on servers (that is, the sum of pre-loaded and flexible servers). We propose an implementable data driven policy that dynamically allocates servers in response to time-varying demand that minimizes the longrun system cost subject to a specified service quality. This policy combines Singular Value Decomposition (SVD) method for forecasting, Stationary Dependent Period by Period (SDPP) paradigm to address the time-varying nature of the system, and simple queueing models with robust Chebyshev bounds for ensuring that service quality constraints are satisfied. We evaluate several competing policies with discrete event simulations using three-year data from the VCL of NC State University and show that our recommended policy achieves the target service quality using less than half of the servers under current policy.