Modeling and Analysis of Availability of Datacenter Power Infrastructure

Realizing highly available datacenter power infrastructure is an extremely expensive proposition with costs more than doubling as we move from three 9’s (Tier-1) to six 9’s (Tier-4) of availability. Existing approaches only consider the cost/availability trade-off for a restricted set of power infrastructure configurations, relying mainly on component redundancy. A number of additional knobs such as centralized vs. distributed component placement, power-feed interconnect topology and component capacity over-provisioning also exist, whose impact has only been studied in limited forms. In this paper, we provide a systematic approach to understand the cost/availability trade-off offered by these configuration parameters as a function of supported IT load. We develop detailed datacenter availability models using Continuous-time Markov Chains and Reliability Block Diagrams to quantify the relative impact of these parameters on availability. Using real-world component availability data to parametrize these models, we offer a number of interesting insights into developing costeffective yet highly available power infrastructure. As two salient examples, we find (i) although centralized UPS placement offers high availability, it does so with significant cost, and (ii) distributed server-level UPS placement is much more cost-effective but does not offer meaningful availability for operating the datacenter at full load. Based on these insights, we propose a novel hybrid strategy that combines the server-level UPS placement with a rack-level UPS, achieving as good availability as existing centralized techniques, at just twothirds of its cost.