Internet service providers face the daunting task of maintaining guaranteed latency requirements while reducing power requirements. In this work, we focus on a class of services with very high cpu and memory demands, best represented by internet search. These services provide strict latency guarantees defined in Service-Level Agreements, yet the clusters need to be flexible to different optimizations, i.e. to minimize power consumption or to maximize resource usage. Unfortunately, standard cluster algorithms, such as resource allocation, are oblivious of the SLA allocations, while power management is typically only driven by cpu demand. We propose a power-aware resource allocation algorithm for the cpu and the memory which is driven by SLA and allows for various dynamic cluster configurations, from energy-optimal to resource-usage-optimal. Using trace-based simulation of two service models, we show that up to 24% energy can be preserved compared to the state-of-art scheme, or maximum memory utility can be achieved with 20% savings.