Optimizing component reliability in datacenters using predictive models

Prior studies have demonstrated component silicon reliability sensitivity to datacenter cooling strategies that include containment and non-containment cooling with return side setpointing. In this work, the evaluations were compared to supply side setpointing to understand the optimal approach to datacenter cooling. This was accomplished through the development of predictive models of T_inlet as a function of T_setpoint. These models were subsequently used for tradeoff analyses to quantitatively evaluate possible ways to leveraging of the improved reliability. Specifically, to reduce data center costs and/or improve performance. Performance tradeoff estimates were based on experimental lab data.