Distributed deep neural network training is widely employed in clusters, which can partition large models and datasets across multiple computation nodes for parallel training. The development of GPUs is far outpacing that of network devices, leading to a communication bottleneck. Existing work focuses on single-iteration optimization, resulting in a large amount of redundant data transmission. We propose GIT, a novel communication optimization scheme for distributed training that filters out similar gradients in adjacent iterations. Our key observation is that some of the gradients between adjacent iterations are similar, and filtering out them can efficiently reduce communication volumes. First, GIT adopts a similarity-based selective transmission strategy. It evaluates the similarity of the gradients of adjacent iterations and filters out the gradients below a preset threshold. Second, GIT proposes a loss-monitoring-based dynamic threshold adjustment strategy, which delicately adjusts the similarity threshold by tracking the trend of training loss to maintain model accuracy. We implement and integrate GIT with PyTorch. In a 100Gbps RDMA testbed, GIT improves the training speed of real-world DNNs by up to 17.7% compared to the alternatives, including OmniReduce, NCCL, and basic PS.