高速互连网络中基于网卡的归约计算硬件卸载机制

聚合通信广泛应用于高性能计算的研究和工程领域。在大规模的科学和工程计算中,聚合通信开销占据很大比例,有时甚至可达到全部消息传递开销的80%,是高性能计算系统的性能瓶颈。因此提出了一种基于网卡的归约计算硬件卸载机制,通过在网卡上嵌入归约操作逻辑部件,实现了数据在传输过程中的计算,减轻了CPU的负担,降低了通信延迟。通过FPGA平台实现了16节点的归约操作实验,并基于xNetSimPlus模拟器模拟了不同节点规模的归约操作。实验证明,卸载机制能有效减少聚合通信中归约操作的时间,所提支持归约计算的网卡卸载机制最高可以加速归约操作2.71倍。

NIC-based offloading mechanism supporting reduction operation on high-speed interconnection system

Collective communication is widely used in the field of high-performance computing research and engineering. In large-scale scientific and engineering computing, collective communication overhead accounts for a large proportion, sometimes even reaching 80% of the total messaging overhead. It is the performance bottleneck of the high-performance computing system. A NIC-based offloading mechanism supporting reduction operation was proposed. By embedding reduction operation logic components on NIC, the calculation of data during transmission was implemented, and the burden on the CPU and the communication delay were reduced. A 16-node protocol operation experiment was realized through the FPGA(field programmable gate array) platform, and the protocol operation in different node size was simulated based on the xNetSimPlus simulator. Experiments show that the method can effectively reduce the time of protocol operation in collective communication, and the proposed NIC offloading mechanism that supports reduction operation hardware offload can accelerate all-reduce operations by up to 2.71 times.