Gazelle Network Optimization

Application scenario: In MySQL heavy-load tests, one of the TPC-C performance bottlenecks is the network protocol stack. The performance of a traditional user-mode protocol stack is poor when the number of cores exceeds 40, which cannot meet the application acceleration requirements of flexible network thread models.

Technical principle: Gazelle introduces the shadow file descriptor (FD) mechanism to implement a cross-NUMA lock-free multi-thread protocol stack with high linearity. It also introduces a cross-thread asynchronous communication mechanism to manage hugepage memory by region and control the memory access distance to realize zero-copy cross-thread data sharing. Through proxying wakeup, Gazelle implements cross-thread communication in a low latency.

Performance metric: The TPC-C comprehensive performance is improved by 10%.