Network Endpoint Congestion Management via Speculative Probing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-performance computing (HPC) networks face challenges in managing endpoint congestion, particularly due to small message traffic, which existing protocols like ECN and SRP struggle to address effectively, leading to slow response times and high overhead.
Innovation Solution
The Last-Hop Reserve Protocol (LHRP) and Small-Message Speculative Reservation Protocol (SMSRP) proactively manage endpoint congestion by using speculative packets that are dropped if congestion is high, generating a negative acknowledgement and reservation to reschedule transmission as non-speculative packets, reducing overhead and latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If ECN protocol is used to signal network congestion, then congestion can be detected and traffic injection rate can be reduced, but response time is slow because ECN is reactive and responds to congestion after it has already occurred
Solution Approach 1:
The patent applies preliminary action by using speculative packets to probe network congestion conditions before actual data transmission. The source endpoint sends small speculative packets ahead of time to detect congestion levels, and based on the response (ACK or NACK), decides whether to proceed with data transmission. This proactive approach prevents congestion before it occurs, eliminating the slow response time inherent in reactive ECN protocols while maintaining reliable congestion detection.
2Reliability
If SRP protocol is used to actively combat endpoint congestion with reservation handshake, then endpoint congestion can be avoided, but latency increases due to reservation overhead and the protocol works well only for large message transfers
Solution Approach 1:
The patent applies partial action by implementing a simplified congestion control mechanism that only performs speculative packet probing and conditional retransmission, without the full reservation handshake overhead of SRP. The source endpoint sends speculative packets and based on the response, either transmits data immediately (if ACK received) or waits and retransmits later (if NACK received). This partial approach provides effective congestion avoidance for small messages while minimizing latency overhead, making it suitable for fine-grained communication workloads.
3Loss of time
If speculative packets are used to mask reservation latency overhead, then latency can be reduced for large messages, but small message traffic or fine-grained communication still causes endpoint congestion
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the speculative packet size and transmission timing based on network conditions and message size. For small messages, the speculative packet size is kept minimal (e.g., 1-8 bytes) to reduce overhead, while for large messages, larger speculative packets can be used to better amortize the probing overhead. The protocol also adjusts the waiting time after receiving NACK before retransmission, optimizing performance for different workload characteristics and message sizes.
Data Source
AI summary
A congestion management protocol that can be used for small messages in which the last-hop switch determines the congestion of the end point. The last-hop switch drops messages when the end point is congested and schedules a retransmission. A second congestion management protocol transmits small messages in a speculative mode to avoid the overhead caused by reservation handshakes.