With the high integration density and complexity of the modern multi-core platform, thermal problems become more and more significant for both the manufacture and system designer. Dynamic thermal management technique is one effective and efficient way to mitigate and avoid thermal emergences. In this paper, we propose a novel predictive dynamic thermal management algorithm to maximize the multi-core system throughput while satisfying the peak temperature constraints. Different from the conventional approaches, we found that it is not necessarily always a good choice to migrate a hot task to the core with the lowest temperature. Instead, in our algorithm, we develop a new temperature prediction technique and migration scheme that take the local temperature of a core as well as the impacts from neighboring cores into considerations. According to our experiment results on a practical Intel desktop platform, the proposed algorithm can significantly improve the throughput compared with the conventional approach.