Host Node Bandwidth Control via Latency-Responsive Token Pool
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In hosted systems, ensuring fair bandwidth allocation between host nodes and data nodes is challenging, as existing protocols like TCP do not effectively prevent a few host nodes from consuming disproportionate amounts of bandwidth, leading to performance issues for other nodes.
Innovation Solution
Implementing a token-based bandwidth control system where each host node manages its bandwidth usage independently by maintaining a token pool, with tokens representing bandwidth usage, and dynamically adjusting the pool size based on latency to throttle data transmission during congestion, ensuring fair allocation without inter-host communication.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If host nodes coordinate write operations with each other to allocate bandwidth fairly, then bandwidth fairness is improved, but communication overhead increases and performance deteriorates
Solution Approach 1:
Each host node independently monitors its own bandwidth usage and adjusts its transmission rate based on locally observed latency metrics. The system eliminates the need for inter-host coordination messages by enabling each node to self-regulate its bandwidth consumption, thereby achieving fair bandwidth allocation without additional communication overhead.
Solution Approach 2:
The system implements a feedback mechanism where each host node monitors latency metrics from its write operations to data nodes. Based on this feedback, the node dynamically adjusts its transmission rate - increasing throughput when latency is low and throttling when latency is high. This closed-loop control enables fair bandwidth sharing without requiring explicit coordination between hosts.
2Reliability
If TCP throttles transmission rate to reduce packet loss, then packet delivery reliability is improved, but bandwidth utilization decreases and host-to-storage performance deteriorates
Solution Approach 1:
The system dynamically adjusts the transmission rate based on real-time latency conditions rather than using static TCP congestion control. When latency is low, the host increases its transmission rate to maximize bandwidth utilization. When latency increases indicating congestion, the host throttles its rate. This dynamic adjustment allows the system to achieve both high bandwidth utilization and reliable packet delivery by adapting to changing network conditions.
3Productivity
If a single host node uses available bandwidth aggressively, then its own write performance is improved, but other host nodes experience degraded performance due to bandwidth starvation
Solution Approach 1:
The system changes the behavioral parameter of each host node from aggressive bandwidth consumption to latency-responsive transmission. Each host monitors latency metrics and adjusts its transmission rate accordingly. When the network is underutilized and latency is low, hosts can transmit aggressively. When congestion occurs and latency rises, hosts automatically reduce their transmission rates, preventing bandwidth starvation and ensuring fair sharing among all hosts.
Data Source
AI summary
Techniques are described that allow each host in a multi-host system to throttle its bandwidth between multiple data nodes without the need to coordinate with the other hosts. Specifically, techniques are described in which a limit is applied to the amount of sent-but-not-acknowledged data a given host may have. If the host has reached the limit, the host must wait for acknowledgements before sending more data. In one embodiment, the limit is enforced using a token-based bandwidth control. Embodiments are described in which the limit on sent-but-not-acknowledged data varies dynamically based on system conditions. Specifically, each host may monitor its aggregate latency, and increase the limit when latency is low (indicating low-congestion conditions), and decrease the limit when latency is high (indicating high-congestion conditions).