In this paper, a runtime performance projection model for dynamic power management is proposed. The model is built as a first-order linear equation using a linear regression model. It could be used to estimate performance impact from different p-states (voltage-frequency pairs). Workload behavior is monitored dynamically for a program region of 100M instructions using hardware performance monitoring counters (PMCs), and performance for the next region is estimated using the proposed model. For each 100M-instructions interval, the performance of all processor p-states is estimated and the lowest frequency is selected within specified performance constraints. The selected frequency is set with a low-overhead DVFS-based (dynamic voltage-frequency scaling) p-state changing mechanism for the next program region. We evaluate the performance degradation and the amount of energy saving of our dynamic power management scheme using the proposed projection model for SPEC CPU2000 benchmark on a Pentium M platform. We measure the execution time and energy consumption for 4 specified constraints - 10%, 20%, 40%, 80%, on the maximum allowed performance degradation. The result shows that our dynamic management scheme saves energy consumption by 3%, 18%, 38% and 48% with a performance degradation of 3%, 19%, 45% and 79% under 10%,20%,40% and 80% constraints, respectively.