DPDSN: data plane deadline-sensitive scheduling in data center networks

Many existing works are focused on minimizing the Flow Completion Time (FCT) for data center networks. For example, D-SRTF performs shortest remaining time first (SRTF) scheduling and determines the remaining time according to the estimated remaining flow size and the available bandwidth. D-SRTF mimics the remaining flow size utilizing bytes sent of each flow. However, the flow size is unknown for many application, using the statistics information of the bytes sent to mimic the remaining time of flow is an approximate estimation which is not accurate; Besides, the flow-level fine-grained scheduling schemes will determine priorities according to different stages in lifetime of the flow, which will lead to larger delay jitter. Moreover, traditional schemes are deployed at the terminal which cannot perceive the network status in time. In this paper, we propose DPDSN, a deadline-sensitive priority-based queue scheduling scheme based on programmable data plane for data center networks. DPDSN implements a distributed scheduling scheme inside the network, which has better awareness of network status; Furthermore, DPDSN performs an urgency-sensitive priority-based queue scheduling mechanism under the deadline. According to the elapsed transmission time of packets, the urgency, which represents the average remaining time under the deadline on each remaining switch along the packet forwarding path, is calculated at each switch. DPDSN achieves a kind of packet-level fine-grained control mechanism, which can greatly reduce the network delay jitter compared with the existing works. Extensive simulations on Mininet show that DPDSN outperforms the currently best implementable scheme which is flow-level fine-grained in terms of average FCT, jitter and deadline miss ratio.