Hybrid Network Flow Scheduling for Data Center Coflow Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data center network management systems for distributed data parallel computing face inefficiencies due to neglect of infrastructure characteristics, leading to suboptimal performance, especially in handling successive computing phases with data dependency, where optimizing the completion time of sets of flows is critical rather than reducing average flow completion time.
Innovation Solution
A dynamic scheduling algorithm and hybrid centralized-decentralized system that collects real-time bandwidth data and manages Coflows with a master-slave architecture, allowing for precise scheduling and bandwidth management without hardware modifications, prioritizing Coflows based on completion time and adjusting transmission rates to optimize network utilization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a centralized system is used to collect and manage Coflows, then deployment is simplified, but real-time bandwidth information accuracy deteriorates due to static configuration
Solution Approach 1:
The system divides bandwidth management into two parts: a centralized controller for deployment simplicity and distributed daemon processes on each switch for real-time bandwidth monitoring. This segmentation allows the centralized system to maintain ease of deployment while distributed components provide accurate real-time measurements.
Solution Approach 2:
Daemon processes are introduced as intermediaries between the centralized controller and the actual network traffic. These daemons run on distributed switches, collect real-time bandwidth information locally, and report to the centralized controller, thereby bridging the gap between centralized management and distributed real-time monitoring.
2Measurement precision
If distributed flow scheduling is implemented, then real-time bandwidth management improves, but hardware switch modifications are required
Solution Approach 1:
The system enables switches to self-configure by running daemon processes that automatically collect bandwidth information and interact with the centralized controller. This eliminates the need for manual hardware modifications while achieving distributed real-time bandwidth management.
Solution Approach 2:
The patent replaces physical hardware modifications with software-based daemon processes that run on existing switches. Instead of modifying the hardware architecture, the system uses software agents to implement distributed bandwidth monitoring and control functions.
3Productivity
If average flow completion time is optimized, then network throughput increases, but coflow completion time deteriorates due to data dependencies
Solution Approach 1:
The scheduling system dynamically adjusts flow priorities based on coflow context and data dependencies. Instead of static priority assignment, the system continuously monitors coflow progress and reconfigures scheduling decisions to balance individual flow throughput with overall coflow completion time.
Solution Approach 2:
The system implements feedback mechanisms where the centralized controller receives real-time status information from distributed daemons about coflow progress and bandwidth utilization. Based on this feedback, the controller dynamically adjusts scheduling decisions to optimize both throughput and coflow completion time.
Data Source
AI summary
The present invention relates to Data Center Network (DCN) flow scheduling scheme. It provides a dynamic scheduling algorithm and a hybrid of centralized and decentralized scheduling system to improve the performance of DCN and data parallel application. The scheduling system uses a central controller to collect the real-time bandwidth of each node, and schedule the priority as well as transmission rate of each network flow set combined by application context (Coflow [1]). The centralized scheduling avoids a sophisticated system design and hardware (switch) modification to comparing with full decentralized solutions. The combination of centralization and decentralization decreases the average completion time of Coflows, and eventually improve the performance of data parallel applications.