With the increasing scale of GPU clusters, resource fragmentation has emerged as a critical issue, leading to prolonged job queuing times and degraded quality of service (QoS). Existing methods such as FGD and Gandiva attempt to alleviate this issue but fall short in addressing fragmentation caused by staggered job completion and lack flexible job migration mechanisms, thereby limiting overall system efficiency. In this paper, we introduce DeFragS, a deep learning (DL) job scheduler that integrates a job placement policy with a job migration policy to mitigate resource fragmentation in GPU clusters. The job placement policy identifies optimal allocation nodes by considering both temporal and spatial factors, thereby reducing fragmentation caused by staggered job completion. Meanwhile, the job migration policy incorporates a scoring mechanism and swap migration, improving the flexibility and effectiveness of the defragmentation process, ultimately leading to near-optimal outcomes. We evaluate DeFragS using production traces from three real-world GPU clusters. DeFragS achieves an up to 8.9% reduction in GPU fragmentation and a 17–45% decrease in average job queuing time across three different clusters, outperforming state-of-the-art methods.