CoFB: latency-constrained co-scheduling of flows and batches for deep learning inference service on the CPU-GPU system

Recent years have witnessed significant achievements in deep learning (DL) technologies. In the meantime, an increasing number of online service operators take advantage of deep learning to provide intelligent and personalized services. Although significant efforts have been put into optimizing the inference efficiency, our investigation shows that for many DL models that process data-intensive requests, the network I/O subsystem also plays an essential role in determining responsiveness. Furthermore, under the latency constraint, uncontrolled network flow processing will impact request batching. Based on the above observation, this paper proposes CoFB, an inference service system that optimizes performance in a holistic way. CoFB improves the load imbalance in the network I/O subsystem with a lightweight flow scheduling scheme that collaborates the network interface card with a dispatcher thread. In addition, CoFB introduces a request reordering and batching policy and an interference-aware concurrent batch throttling strategy for enforcing inference concerning the deadline. We evaluate CoFB on four DL inference services and compare it to two state-of-the-art inference systems: NVIDIA Triton and DVABatch. Experimental results show that CoFB outperforms these two baselines by serving up to 2.69x and 1.96x higher load under preset tail latency objectives, respectively.