Dynamic multi-tiered Web system can be affected by many unexpected factors during runtime. At the same time, it has different performance characteristic under different workload pattern condition. Accordingly, different performance models are needed to describe the system. Currently, performance guarantee mechanism based on feedback control theory, which aims to eliminate the effect of unexpected factors, often employs single and fixed system performance model which can not adapt to changing performance characteristics. Under fluctuant and unexpected workload condition of Internet, the method can decrease the precision and stability of performance. To guarantee anticipated response time goal, this paper presents an adaptive performance guarantee mechanism based on online estimation of performance model. Using two complementary transactional Web benchmarks to evaluate the techniques experimentally shows that the mechanism can efficiently reduce the squared departure of performance goal under varying workload.