CRSP: Network Congestion Control through Credit Reservation

In HPC, endpoint congestion can lead to the generation of saturated trees, which seriously affect the performance of network systems. Endpoint congestion can be mitigated by reducing the injection rate of the source, but requires fast reaction time to avoid the formation of congestion. In remote direct memory access (RDMA) network interface card (NIC) enabled with offload functions, tasks are transmitted in the form of flow for applications. Therefore, flow completion time and throughput are important factors that determine the efficiency of the application execution. However, the existing hardware solutions do not take into consideration of the impact of flow completion time on the applications. For a time-division reservation protocol, such as speculative reservation protocol (SRP), multiple flows in the source may conflict due to the overlap of the grant time. The reservation operation is inaccurate as a result. Therefore, we present CRSP, a novel, low-cost credit-reserved scheduling protocol to efficiently avoid the endpoint congestion in HPC. The CRSP avoids the formation of endpoint congestion by making reservations to the credit of the destination, while also minimizing the overhead caused by the reservation operation. We evaluate our CRSP protocol against the reservation based protocol, SRP, and the simulation results show that the average flow latency can be reduced by 29.67% in hot spot mode compared with SRP.