RDMA Queue Prioritization in NICs for Low-Latency Service Traffic

The patent applies local quality by assigning different service levels to different RDMA requests based on their importance. High-service-level requests are prioritized over low-service-level requests, creating differentiated processing quality for different types of requests. This resolves the contradiction by maintaining simple FIFO processing for most requests while providing accelerated processing for important requests through service level tagging and priority-based queue selection.

The patent implements preliminary action by pre-tagging RDMA requests with service level indicators before they enter the processing queue. The network interface card pre-identifies high-priority requests and places them in appropriate priority queues, so that when processing occurs, the prioritization is already established. This reduces processing latency for important requests without adding complex real-time decision-making during the processing phase.

The patent segments queue context storage into two parts: frequently accessed context information is stored in the network interface card's onboard memory (fast but limited capacity), while less frequently accessed or larger context data remains in host memory (slow but large capacity). This segmentation allows the system to achieve fast access speeds for critical operations while maintaining sufficient storage capacity through the host memory backend.

The network interface card's onboard memory acts as an intermediary between the processor and host memory for queue context access. Instead of directly accessing host memory for every queue context operation, the NIC caches relevant context in its own memory, providing fast access while the host memory serves as the ultimate storage backend. This intermediary approach resolves the speed-capacity contradiction by providing fast access for frequent operations while maintaining large storage capacity.