Network Congestion Detection and Abatement in Parallel Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Massively parallel processing systems face network congestion issues due to buffer overflow, leading to delays in packet delivery, which negatively impacts the execution speed of jobs across nodes, especially when many nodes send requests to a single destination node.
Innovation Solution
A congestion detection system comprising a detection, abatement, and causation system that monitors network components, limits packet injection rates, and identifies the source of congestion by analyzing performance metrics and ejection rates to alleviate network congestion and pinpoint the cause.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If many nodes send requests to a single destination node, then the computational capacity and network utilization increase, but buffer overflow occurs leading to packet delivery delays
Solution Approach 1:
The system performs preliminary congestion detection by monitoring buffer occupancy levels before buffers become completely full. When congestion is detected (buffer occupancy exceeds a threshold), the system proactively abates congestion by limiting packet injection rates before buffer overflow occurs, preventing packet delivery delays rather than reacting after delays have occurred.
2Productivity
If packet injection rate is increased to improve job execution speed, then productivity increases, but network congestion and buffer overflow occur
Solution Approach 1:
The system implements a feedback mechanism where congestion detection monitors buffer occupancy levels in real-time and provides feedback to the congestion abatement mechanism. When congestion is detected, the system adjusts packet injection rates by limiting traffic from specific sources, creating a closed-loop control system that dynamically balances productivity and congestion prevention based on current network conditions.
3Reliability
If congestion detection and abatement mechanisms are implemented, then packet delivery reliability improves, but system complexity increases
Solution Approach 1:
The congestion detection and abatement system operates autonomously without requiring external intervention. The congestion detection mechanism automatically monitors buffer levels, identifies congestion conditions, and triggers abatement actions by limiting packet injection from identified sources. The system self-manages the complexity of congestion control, providing reliable packet delivery while minimizing the need for external configuration or management overhead.
Data Source
AI summary
A method and system for detecting congestion in a network of nodes, abating the network congestion, and identifying the cause of the network congestion is provided. A congestion detection system may comprise a detection system, an abatement system, and a causation system. The detection system monitors the performance of network components such as the network interface controllers and tiles of routers to determine whether the network is congested such that a delay in delivering packets becomes unacceptable. Upon detecting that the network is congested, an abatement system abates the congestion by limiting the rate at which packets are injected into the network from the nodes. Upon detecting that the network is congested, a causation system may identify the job that is executing on a node that is the cause of the network congestion.