Congestion Root Detection in Data Center Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing congestion management solutions in data center networks often reduce data rates for all flows, including innocent ones, due to the propagation of congestion trees, leading to data packet loss and increased delay.
Innovation Solution
The technique identifies and isolates the root of congestion within the network, using specific Q-Set parameters and flow control signals to selectively control congestion, thereby avoiding data packet loss and minimizing delay for non-contributing flows.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If link level flow control is used to manage congestion, then congestion can be controlled at the link level, but congestion trees are created and expanded that cause data rate reduction for innocent flows
Solution Approach 1:
The patent segments the network flows into two categories: contributing flows (those causing congestion) and non-contributing flows (innocent flows). By identifying the root of congestion and tracing only the contributing flows, the system applies flow control selectively rather than globally, thus preventing the propagation of congestion trees to innocent flows while still managing congestion effectively at the root.
Solution Approach 2:
The patent applies different flow control actions to different flows based on their contribution to congestion. At the congestion root, the system identifies which specific flows are contributing to the congestion and applies rate reduction only to those flows, while leaving non-contributing flows unaffected. This localized approach ensures that flow control effectiveness is maintained for problematic flows while preserving throughput for innocent flows.
2Reliability
If network level congestion management reduces traffic rate at the source, then congestion at network hot spots is controlled in the long term, but data packet loss and increased delay occur in the short term before control takes effect
Solution Approach 1:
The patent implements preliminary action by detecting congestion at the root and immediately identifying contributing flows before the congestion propagates extensively through the network. By taking immediate action at the congestion root to reduce rates of contributing flows, the system prevents the formation and expansion of congestion trees, thereby avoiding the short-term delays and packet losses that would otherwise occur while network-level control takes effect.
Solution Approach 2:
The patent introduces an intermediary mechanism that operates between link-level flow control and traditional network-level congestion management. This intermediary identifies the root of congestion and the specific contributing flows, then applies targeted rate reduction to these flows. This intermediate approach bridges the gap by providing faster response than traditional network-level management while avoiding the excessive packet loss and delay problems of untargeted link-level control.
3Reliability
If congestion notification is performed through packet marking, then sources are notified of congestion, but all flows on the marked path are affected including innocent flows
Solution Approach 1:
The patent segments the notification mechanism to target only contributing flows rather than all flows on a path. By identifying the root of congestion and tracing which specific flows are contributing to it, the system sends congestion notifications selectively to sources of contributing flows. This segmentation ensures that innocent flows are not notified and thus not unnecessarily rate-reduced, while still providing effective congestion notification to the flows that actually caused the congestion.
Solution Approach 2:
The patent applies congestion notification locally and selectively rather than globally along entire paths. At the congestion root, the system identifies which specific flows are contributing to the congestion and notifies only those flows' sources. This localized notification approach ensures that congestion information is communicated effectively to the relevant flows while leaving innocent flows unaffected, thus maintaining their data rates.
Data Source
AI summary
A device for performing congestion detection and management at a node of a network may include a congestion management module to monitor an arrival and a departure rate of data packets associated with a queue, a queue size, or a rate of change of the queue size. The congestion management module may identify the queue as a congested queue by one of determining that the arrival rate of the data packets associated with the queue is larger than the departure rate of the data packets associated with the queue, the queue size is larger than a first threshold, or the rate of change of the queue size is larger than a second threshold. The congestion management module may identify the congested queue as a congestion root by determining that the congested queue is not affected by a flow-control signal. A queue buffer temporarily stores data packets associated with the queue.