RoCE Switch Signaling for Selective Queue Congestion Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing RDMA over Converged Ethernet (RoCE) protocol faces challenges in effectively controlling network congestion, leading to service delay and failure in large-scale high-concurrency scenarios, particularly due to inefficiencies in congestion notification and control mechanisms like DCQCN.
Innovation Solution
A network congestion control method that involves switches exchanging signaling packets with flow source information to implement priority-based flow control (PFC) packets, allowing switches to manage queue backlogs and ensure low service delay without affecting throughput, by distinguishing between different congestion states and sending appropriate PFC packets to source servers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If DCQCN congestion control mechanism is used in large-scale high-concurrency scenarios, then network throughput is maintained, but queue backlog increases and service delay occurs
Solution Approach 1:
The patent segments the network flow control into multiple priority levels (PFC0-PFC7) corresponding to different queue backlogs. When congestion is detected at specific thresholds, PFC packets are sent to suspend transmission in corresponding priority queues, preventing end-to-end backlog while maintaining throughput for non-congested flows
Solution Approach 2:
The patent applies local quality by sending PFC flow control packets selectively to specific network devices based on where congestion is detected. Instead of global flow control, only the congested link's source device receives PFC packets, allowing local suppression of queue backlog without affecting other network paths
2Loss of time
If flow control is implemented to suppress queue backlog, then service delay is reduced, but network throughput may be affected
Solution Approach 1:
The patent applies partial action by implementing flow control only on specific queues experiencing congestion rather than all queues. PFC packets are sent selectively to suspend transmission in affected priority queues while leaving other queues unaffected, thus suppressing backlog locally without reducing overall network throughput
3Device complexity
If conventional TCP/IP protocol is used, then network management is simplified, but CPU overhead becomes extremely high
Solution Approach 1:
The patent implements self-service by enabling network devices to autonomously detect congestion conditions and send PFC packets without requiring complex TCP/IP stack processing. The congestion control mechanism operates at the data link layer, eliminating the need for kernel intervention and high-CPU operations associated with conventional TCP/IP protocols
4Use of energy by moving object
If RDMA protocol is used to reduce CPU overhead, then processing efficiency improves, but network congestion control becomes ineffective
Solution Approach 1:
The patent introduces PFC packets as an intermediary mechanism between RDMA network devices. These control packets carry congestion information and flow control instructions, enabling effective congestion management in RDMA networks without requiring CPU intervention in the data path, thus maintaining both low CPU overhead and effective congestion control
Data Source
AI summary
This application discloses a network congestion control method, apparatus, device, and system, and a storage medium. In an example method, a first switch receives a target signaling packet that is sent by a second switch in a case in which a network congestion status is a target network congestion status. The target signaling packet carries flow source information. The first switch sends, based on the target signaling packet, target flow control information to a network device corresponding to the flow source information. The target flow control information is used to indicate flow control.