CALM: Survivable Virtual Data Center Allocation in Cloud Networks

Abstract

Cloud data centers have become a popular infrastructure to host diversified application services for tenants. To provide agility and elasticity in resource usage for cloud services, the virtual data center (VDC) is proposed to allocate both virtual machines (VM) and network bandwidth. However, at cloud scale, hardware (e.g., link, server, and switch) failures are inevitable, which may lead to degradation in service performance. To address this challenge, we study the survivable virtual data center allocation problem (SVAP), which aims at allocating survivable virtual data center (SVDC) to each tenant to guarantee resource demands will always be satisfied even after failures. Our objective is to minimize the total bandwidth consumption in order to accommodate more SVDCs. We prove that SVAP is NP-hard and design the Collocation-Aware survivable VM placement and Link Mapping algorithm (CALM). CALM solves the problem in two stages, i.e., VM placement (VMP) and virtual link mapping (VLM). We further find that without an appropriate VMP strategy, VLM cannot lead to the minimum network resource usage. Therefore, we propose a polynomial-time algorithm called collocation-aware survivable placement (CASP) for VMP. For the VLM stage, we formulate a linear programming model to map flows onto the data center network in order to ensure survivability under switch failures. We evaluate the performance via simulations and show that CALM could save up to 42 percent network resource compared to the baseline algorithm. We further show that CALM uses only additional 13 percent network resource to guarantee survivability as compared to a typical VDC strategy.