Advanced cooling concepts and workload optimization for data centers
The successful operation of a large-scale air-cooled data center (DC) requires an efficient cooling system to ensure that the DC operator can provide its services to customers with maximum availability and reliability at minimal operational cost. An efficient cooling system guarantees that the temperatures at the inlets of all computing devices in the DC racks never exceed a given threshold value. However, most of today's DCs waste cooling energy because they are operated at significantly lower temperatures than actually necessary. This approach certainly reduces the potential risk of reacting too late to a harmful temperature increase in the DC, but can lead to the consumption of 4 to 6% more cooling energy for each degree Celsius below the upper temperature limit.
To realize these potential savings in cooling energy, we are designing advanced sensing and control concepts to optimize the air flows and temperature distribution in a DC room. The results demonstrate that up to 20% of cooling energy can be saved by applying dynamic air-flow control. Using these control concepts in a large-scale DC and incorporating the control strategy into a WSAN (wireless sensor and actuator network) are issues we will study next.
Equally important is the correct placement of workloads onto server clusters in data centers. With the advances in virtualization technologies, workloads are mobile and can be executed on different physical servers in the in-house or in geographically dispersed data centers or even by a cloud provider. In the first two cases the issue arises concerning the optimal placement with respect to SLA attainment and—energy—cost, whereas in the case of a cloud provider the issue is only the SLA attainment. Consolidation of multiple workloads on the same physical server requires careful planning and workload estimation to prevent unexpected degradation of individual application performance. The increasing trend of dynamic workload consolidation on heterogeneous hardware further exacerbates challenges in data center workload management.
We aim to achieve maximum resource usage while respecting individual application performance targets and energy constraints. In addition, in clustered systems we work on optimal workload balancing technologies, which work in an entirely non-intrusive manner and can be deployed on a wide range of systems. Ultimately we want to couple the above-mentioned sensor network information to determine the best possible location for a workload to be executed, thereby achieving minimal energy consumption with little or no impact on SLA attainment.