The demand for large-scale deep learning is increasing, and distributed training is the current mainstream solution. Decentralized algorithms are widely used in distributed training. However, in a heterogeneous environment, each worker computes the same amount of training data resulting in a lot of wasted time for waiting the straggler. In this paper, we proposed a self-adaptive task allocation algorithm (SATA) which allows that each worker acquires the amount of training data adaptively based on the performance of workers in the heterogeneous environment. In order to show the applicability of SATA in heterogeneous clusters better, we set up the heterogeneous cluster composed of two or three different types of GPUs. Besides, we conduct a series of experiments to show the performance of SATA. The experimental results illustrate several advantages of SATA. SATA can accelerate distributed training about 3.3X that of All-Reduce as well as about 1.9X-3.8X that of AD-PSGD algorithm. And the total training time of SATA is reduced from 20 to 40 percentage compared to All-Reduce. © 2022 Knowledge Systems Institute Graduate School. All rights reserved.