CPU efficient networking stack using NIC memory

Abstract

CPU speed has stalled while network speed has been growing exponentially. This imbalance of speed improvements makes the host CPU the bottleneck. To alleviate it, many studies in the network field have been conducted to reduce the CPU burden. We cluster the previous works into the following three main categories based on the approaches: (1) to reduce CPU usage by implementing a networking stack on NIC, (2) to eliminate payload copy, and (3) to optimize the networking stack. Each method, however, has its own drawbacks. (1) The hardware-implemented networking stacks leave little room for modification. (2) Many zero-copy methods require application modification. Further, the application must ensure that the application buffer is not modified for a long time. (3) Copy operation still remains regardless of networking stack optimizations. In this work, we redesign the networking stack to keep the socket buffer in the NIC memory while keeping the socket APIs intact to overcome the mentioned limitations. (1) This work performs the network protocol processing based on the CPU to maintain development flexibility. (2) The application quickly reuses its own buffer, since the socket buffer exists in the NIC memory, unlike the existing zero-copy methods. Further, our work does not require application modification since we keep the socket APIs intact. (3) The NIC DMA module transfers the application buffer to NIC memory without using the host CPU, eliminating CPU-based copy. Our I/O method reduces CPU usage by up to 70.2% compared to the CPU-based memory copy I/O scheme. The TCP stack implemented on top of our work reduces the application buffer holding time by 1.7x $\sim$ 4.9x depending on the buffer size compared to the user buffer direct I/O scheme.