HPC Packet Retransmission for Link, Node, and Congestion Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing high-performance computing systems lack the ability to distinguish between different types of packet transmission issues such as link failures, destination node congestion, and node failures, leading to inefficient and unreliable data exchange.
Innovation Solution
Implementing a node configuration in high-performance computers that utilizes multiple timeouts and heartbeat messages to detect and manage specific packet transmission issues, including a first timeout for link failures, a second timeout for node down scenarios, and a third timeout for congestion, allowing for differentiated handling of each issue.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single timeout mechanism is used to detect packet transmission issues, then the system is simple to implement, but it cannot distinguish between different types of network issues (link failure, congestion, node down)
Solution Approach 1:
The patent segments the single timeout mechanism into three distinct timeout mechanisms with different time periods: first timeout for link failure detection, second timeout for node down detection via heartbeat messages, and third timeout for congestion detection via negative acknowledgments. Each timeout is tailored to detect specific network issues, enabling precise differentiation between link failures, node failures, and congestion conditions.
Solution Approach 2:
The patent introduces intermediary messages to facilitate precise issue detection: heartbeat messages are sent periodically to detect node availability, and negative acknowledgment messages are sent by congested nodes to indicate congestion status. These intermediary messages act as mediators between the sender and receiver, enabling the sender to accurately determine the state of the network and the receiver without direct observation.
2Reliability
If periodic heartbeat messages are sent to all nodes to detect failures, then node availability can be detected, but the network generates a lot of messages in large clusters
Solution Approach 1:
The patent applies local quality by making heartbeat message transmission selective rather than universal. Heartbeat messages are sent only to nodes with which the current node has pending transactions, not to all nodes in the network. This localized approach maintains reliable detection of node availability for relevant communications while significantly reducing the total number of heartbeat messages generated in large clusters.
3Productivity
If packets are retransmitted immediately upon timeout, then data exchange speed is maintained, but congestion may worsen when destination nodes are overloaded
Solution Approach 1:
The patent implements dynamic retransmission behavior based on network conditions. When a sender receives a negative acknowledgment from a congested destination node, it adjusts its retransmission timing by waiting for a predefined time period before retrying. This dynamic adjustment allows the system to maintain productivity by eventually retransmitting packets while preserving reliability by avoiding immediate retransmission that would worsen congestion.
Solution Approach 2:
The patent establishes a feedback mechanism where congested destination nodes send negative acknowledgment messages back to senders to indicate their congested state. Senders use this feedback information to adjust their retransmission behavior, waiting for a predefined time period before retrying packet transmission. This feedback loop enables the system to respond adaptively to congestion conditions, maintaining both productivity and network stability.
Data Source
Figure 1~2
AI summary
The invention concerns a High-Performance Computer (1) comprising a plurality of nodes (10) wherein each node is configured to: - Wait for a first time period after sending a packet, and retransmit said packet if an acknowledgement for the packet has not been received in the first time period, - Periodically send a heartbeat message to each other node having a pending transaction with said node, wait for a second time period after sending a heartbeat message, abort the pending transaction if an acknowledgement for the heartbeat message has not been received in the second period, - When receiving a negative acknowledgement after sending a packet, retransmit said packet after a third time period starting from the reception of the negative acknowledgement, - Wait for a fourth time period and drop a packet if said node is congested during the fourth period and send a negative acknowledgement message.