RoCE Acknowledgment Feedback for Faster Congestion Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing RoCE network congestion control methods suffer from delayed congestion notification and inability to promptly adjust packet sending rates due to the limitations of independent CNP packets, leading to inefficient bandwidth utilization.
Innovation Solution
Implementing a method that uses acknowledgment packets with integrated indication and congestion information, such as a Congestion Extended Transport Header (CETH), to provide real-time network status and enable faster congestion control adjustments, including adjusting sending rates based on congestion degree, location, and queue depth.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If independent CNP packets are used for congestion notification, then the RoCE protocol can notify network congestion to the source end, but the congestion notification is delayed due to the upper limit of packet sending rate at the receive end
Solution Approach 1:
The patent combines congestion notification information with acknowledgment packets into a single unified packet structure. The acknowledgment packet now carries both acknowledgment data and congestion notification data, eliminating the need for separate CNP packets. This merging reduces the total number of packets that need to be sent and received, thereby reducing notification delay while maintaining reliable congestion notification delivery.
2Loss of information
If independent CNP packets are used for congestion notification, then the source end can be notified of network congestion, but the source end cannot be notified when network congestion is eliminated
Solution Approach 1:
The acknowledgment packet is designed to serve multiple functions simultaneously: it provides acknowledgment for received packets, notifies congestion status when congestion occurs, and notifies when congestion is eliminated. This multi-functional design ensures complete congestion status information is transmitted without requiring separate mechanisms for each notification type, thus reducing information loss without significantly increasing system complexity.
3Reliability
If the source end periodically detects network congestion elimination, then it can determine when congestion is cleared, but the packet sending rate cannot be restored in time, affecting network bandwidth utilization
Solution Approach 1:
The patent implements a feedback mechanism where the receive end actively notifies the source end of congestion elimination status through the acknowledgment packet. When the receive end detects that network congestion has been eliminated, it immediately includes this information in the acknowledgment packet sent back to the source end. This active feedback allows the source end to restore its packet sending rate immediately upon receiving the notification, rather than waiting for periodic detection, thereby significantly improving network bandwidth utilization while maintaining reliable congestion elimination detection.
Data Source
AI summary
An RoCE network congestion control method includes a first network device that sends a RoCE protocol packet to a second network device. The first network device receives an acknowledgment packet from the second network device that includes indication information and acknowledgment information in response to the ROCE protocol packet. The indication information indicates whether a network path between the first network device and the second network device is congested. The first network device performs congestion control based on the acknowledgment packet.