Data Center Network Path Verification for Congestion Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Data center networks face unpredictable performance due to incast traffic, congestion, and failure of switches or links, making it difficult to predict throughput and locate performance issues.
Innovation Solution
The method involves sending packets with varied header fields to balance load across multiple paths, using acknowledgment packets to measure latency and loss on each path, and controlling traffic rate at the receiver to avoid congestion, while encapsulating legacy protocols to monitor network health without disrupting them.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If switches use hash function to route packets to single path, then packet order is maintained, but network congestion occurs on specific links
Solution Approach 1:
The patent dynamically changes the hash function input by modifying packet header fields (such as TCP sequence numbers or IP identification fields) to vary the selected path for each packet or flow. This dynamic adjustment allows the system to adapt to network conditions and distribute traffic more evenly across multiple paths, preventing congestion on any single link while maintaining packet ordering through sequence tracking.
Solution Approach 2:
The invention changes parameters in the packet header (such as TTL values, protocol identifiers, or custom fields) to influence path selection. By modifying these parameters, the system can direct different packets or flows through different paths in the network, balancing load across available routes and improving overall network throughput while still ensuring packets arrive in order through reassembly mechanisms.
2Measurement precision
If probe test traffic is sent at high rate to detect failures, then network problems are revealed, but operational traffic is disrupted
Solution Approach 1:
The patent merges probe traffic with operational data traffic by embedding probe packets within the same network flows as regular data packets. This integration allows the system to monitor network performance and detect failures using existing traffic patterns without requiring separate high-rate probe traffic that would disrupt operational flows. The probe information is extracted from normal packet sequences.
Solution Approach 2:
The system uses its own operational traffic to perform self-diagnosis and performance monitoring. By analyzing delays, packet loss, and path variations in its normal data flows, the network can detect failures and performance issues without external probe traffic. This self-service approach eliminates the need for disruptive separate testing while maintaining accurate failure detection.
3Productivity
If TCP fills queue to capacity to sense network capacity, then network utilization is maximized, but latency increases for other packets
Solution Approach 1:
The patent implements feedback mechanisms where receivers send acknowledgment packets that provide information about network conditions, queue status, and path quality back to senders. This feedback allows senders to adjust their transmission rates dynamically, filling queues to optimal levels without overfilling, thereby maximizing network utilization while controlling latency. The feedback loop enables coordinated rate adjustment across the network.
Solution Approach 2:
The system uses periodic acknowledgment packets and flow control signals to regulate traffic rates. Instead of continuously filling queues to maximum capacity, the network employs periodic feedback intervals to adjust transmission rates, allowing queues to be filled to appropriate levels without causing excessive latency. This periodic control mechanism balances utilization and delay.
Data Source
AI summary
Packets in a data communications network are encapsulated by an encapsulation module on a sending computer and decapsulated on the receiver computer, the transmission of data packets being controlled by credit sent by the receiving computer to avoid causing congestion. The encapsulation module varies fields in the packets that are used by switches to determine the path to the destination, so as to distribute the load of a transfer across a plurality of paths to the receiving computer. The sending and receiving computers use per path packet delivery, loss, latency and packet trimming information to detect abnormal network behavior and submit alerts and summary statistics to a monitoring station. The monitoring station uses this information to detect network bottlenecks and other faults and to localize them to specific switches or links.