Prefetching for a shared direct memory access (DMA) engine

A system is disclosed for fetching control instructions for a direct memory access (DMA) engine shared between a plurality of threads. For a data transfer from a first thread by a DMA engine, the DMA engine fetches and processes a predetermined number of control instructions (or work queue elements) for the data transfer, each of the control instructions including an amount and location of data to transfer. The DMA engine determines a total amount of data transferred as a result of the data transfer. The DMA engine then determines a difference between the total amount of data transferred and a threshold amount of data, wherein the threshold amount of data indicates a preferred amount of data to be transferred for the first thread. The predetermined number of control instructions to fetch is updated based on the determined difference.