Network Interface Controller Queue Scheduling for Latency Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional network interface controllers (NICs) face scalability issues with multiple data streams of the same traffic class due to large gate control lists and increased latency, especially when handling periodic data streams with different reduction ratios.
Innovation Solution
The proposed solution involves mapping all data streams corresponding to a traffic class to a single queue and scheduling them based on reduction ratios rather than traditional linear scheduling. This approach reduces the number of queues and gate control list entries, thereby decreasing the area required and latency. Additionally, the NIC is designed to prepare for transmission without waiting for data availability, allowing for more up-to-date payload transmission through interrupt-driven mechanisms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional linear scheduling with separate queues is used for each data stream, then each data stream can be managed independently, but the number of queues and gate control list entries increases, leading to larger device area and increased latency
Solution Approach 1:
The patent merges multiple data streams of the same traffic class into a single queue instead of maintaining separate queues for each stream. The gate control list is also consolidated to manage multiple streams within the same queue, reducing the total number of entries needed. This combining approach maintains the ability to manage different data streams independently through software configuration while significantly reducing hardware complexity and device area.
2Measurement precision
If the NIC waits for data availability before preparing transmission, then data accuracy is maintained, but end-to-end latency increases
Solution Approach 1:
The NIC performs preliminary actions by preparing transmission frames and allocating resources before the actual payload data is available. It creates placeholder frames with metadata and initiates transmission preparation in advance, then updates the frames with the actual payload data when it becomes available. This preliminary action reduces end-to-end latency while maintaining data accuracy through the interrupt-driven update mechanism.
3Adaptability or versatility
If multiple periodic data streams with different reduction ratios are handled, then scheduling flexibility is improved, but the gate control list size increases significantly
Solution Approach 1:
The patent implements a universal gate control list entry format that can handle multiple periodic data streams with different reduction ratios using a single queue. The gate control list entries are designed to be multi-functional, supporting various scheduling scenarios (different reduction ratios, time slots, and traffic classes) within a unified structure. This universality allows the system to maintain high scheduling flexibility while minimizing the number of gate control list entries required.
Data Source
AI summary
Methods, apparatus, systems, and articles of manufacture to transmit and/or receive data streams with a network interface controller are disclosed. An example apparatus includes a direct memory access engine to fetch a descriptor for a data transmission from system memory; and determine a time to generate an interrupt based on the descriptor; a scheduler to trigger the interrupt when the time occurs, the interrupt to cause an application to sample data and store the sampled data as a payload data structure into the system memory; the direct memory access engine to access the payload data structure from the system memory; and the scheduler to cause transmission of the payload data structure.