Per-destination max-min fairness for batch completion time
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In many-to-one communication patterns in data center applications, the completion time is often dictated by the slowest worker node, leading to inefficiencies in bandwidth usage and uneven transmission times across the network.
Innovation Solution
A per-destination max-min fairness scheme is implemented, which iteratively identifies nodes with additional capacity and increases bandwidth allocation to virtual links, ensuring uniform transmission times and maximizing throughput by distributing data flows across multiple paths.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the server waits for responses from all worker nodes before proceeding, then data accuracy is ensured, but the total completion time is dictated by the slowest worker node
Solution Approach 1:
The patent segments the network into multiple layers (first layer with source nodes, second layer with intermediate nodes, third layer with destination nodes) and implements routing decisions at each layer. This segmentation allows parallel processing of routing decisions and enables the system to proceed with aggregated results rather than waiting for all individual worker node responses, thus reducing completion time while maintaining data accuracy through proper aggregation.
Solution Approach 2:
The patent performs preliminary routing decisions and bandwidth allocation before data transmission begins. The system pre-calculates optimal paths through the multi-layer network and allocates bandwidth accordingly, allowing data to flow efficiently without delays during transmission. This preliminary action eliminates waiting time while ensuring data reaches destinations accurately.
2Productivity
If bandwidth is allocated to all worker nodes, then throughput is maximized, but bandwidth utilization becomes uneven across the network
Solution Approach 1:
The patent implements local quality by allocating different bandwidth proportions to different paths based on their specific characteristics and current load conditions. Rather than uniform allocation, each link in the multi-layer network receives bandwidth proportional to its capacity and demand, optimizing overall utilization while maintaining high throughput. This is achieved through the routing determination that considers local network conditions at each layer.
Solution Approach 2:
The system dynamically adjusts bandwidth allocation parameters based on network conditions and flow requirements. By changing bandwidth allocation parameters adaptively rather than using fixed allocation, the system maximizes throughput while avoiding waste on oversubscribed paths and ensuring adequate bandwidth for underutilized paths.
3Productivity
If data flows are routed through multiple paths, then bandwidth utilization increases, but routing complexity increases
Solution Approach 1:
The patent reduces routing complexity by segmenting the network into three distinct layers with clear roles. Source nodes in the first layer send data to intermediate nodes in the second layer, which then forward to destination nodes in the third layer. This segmentation simplifies routing decisions at each layer compared to managing complex end-to-end paths, while still enabling multiple paths and high bandwidth utilization through the layered structure.
Solution Approach 2:
The intermediate nodes in the second layer serve multiple functions: they receive data from multiple source nodes, perform routing decisions, allocate bandwidth, and forward to multiple destination nodes. This multi-functionality reduces overall routing complexity by consolidating routing intelligence at the intermediate layer rather than requiring complex routing logic at every node.
Data Source
AI summary
The present disclosure describes a system and method for reducing total batch completion time using a per-destination max-min fairness scheme. In a distributed computer system, worker nodes often simultaneously return responses to a server node. In some distributed computer systems, multiple batches can traverse a network at any one given time. The nodes of the network are often unaware of the batches other nodes are sending through the network. Accordingly, in some implementations, the different batches encounter different effective path capacities as nodes send flows through links that are or become bottlenecked. The per-destination max-min fairness scheme described herein reduces the total batch completion time by collectively routing the batches in a way that targets providing substantially uniform transmission times without under underutilizing the network.