Latent Dirichlet Allocation (LDA) is a statistical approach for topic modeling with a wide range of applications. In spite of the significance, we observe very few attempts from system track to improve LDA, let alone the algorithm and system codesigned efforts. To this end, we propose eLDA with an algorithm-system codesigned optimization. Particularly, we introduce a novel three-branch sampling mechanism to taking advantage of the convergence heterogeneity of various tokens in order to reduce redundant sampling task. Our evaluation shows that eLDA outperforms the state-of-the-arts.