A server enclosure hosts multiple servers together with cooling fans and needs to control their total power consumption under a given budget. Existing methods for computer system power control either only manipulate CPU power states or assume all the workloads have homogeneous optimization goals, leaving workload-specific optimization with different hardware actuators an unsolved problem, especially at the enclosure level. In this paper, we propose EncloPC, a hierarchical framework for enclosure-level power control. During a power peak, EncloPC dynamically determines the power allocation between the servers running interactive and batch workloads, respectively. EncloPC features two power controllers for the two groups of servers, designed rigorously based on feedback control theory to guarantee the power control accuracy even with modeling errors. Both controllers jointly adapt CPU and memory frequencies for maximized power efficiency and profits. The evaluation results show that compared to state-of-the-art solutions, EncloPC can achieve up to 70% reduction in processing delay for interactive workloads, and an additional annual profit of up to $6.7 Million for a data center hosting 50,000 servers.