Abstract: Disclosed are a data caching method and apparatus for multiple concurrent deep learning training tasks. The method includes: step 1, executing preheating training for each task, collecting feature parameters of training batch samples, and sorting all the tasks according to the collected feature parameters; step 2, calculating the sample number of each training batch hit in a cache of each task under system pre-allocation, and the expected sample number of each training batch hit in the cache of each task; step 3, concurrently executing deep learning training by using a cache dynamic allocation and management strategy; and step 4, when each task enters a last training epoch, adding no new sample data to the caches of these tasks, gradually releasing the occupied cache, and making the released cache to be used by other tasks that are not finished.