Network Link Power Optimization via Communication Pattern Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High performance computing (HPC) systems face inefficiencies in power usage as networking components consume increasing amounts of power, often leading to under-provisioned power budgets, especially as cluster sizes grow, and existing methods fail to optimize network power usage effectively by not leveraging application-specific communication patterns.
Innovation Solution
The method involves analyzing a distributed computer program to estimate virtual network communication traffic, mapping it to physical network links, identifying communication patterns, and categorizing links based on intensity, allowing for dynamic power optimization by turning off or placing unused links in low-power mode, using middleware libraries like MPI to inform network switches about link usage and adjust routing accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If network links are kept active through synchronization messages and explicit configuration to maintain performance, then network performance is improved, but power consumption increases
Solution Approach 1:
The patent implements dynamic link management where network links are actively managed based on real-time communication patterns. The system transitions links between active and low-power states dynamically, using synchronization messages to coordinate these transitions while maintaining performance for critical links and reducing power consumption for non-critical links.
Solution Approach 2:
The patent applies different power management strategies to different network links based on their specific communication patterns and performance criticality. By analyzing application-specific communication patterns, the system identifies which links require high performance and which can tolerate reduced performance or power-saving modes, applying local optimization rather than uniform power management across the entire network.
2Productivity
If cluster size is increased to improve computing capacity, then productivity is improved, but power budget becomes under-provisioned
Solution Approach 1:
The patent changes the power consumption parameter of network links dynamically based on communication patterns. By monitoring and analyzing application communication behaviors, the system adjusts power allocation to match actual usage patterns, enabling larger clusters to operate within existing power budgets by reducing power consumption where full performance is not required.
Solution Approach 2:
The patent performs preliminary analysis of application communication patterns to predict which network links will be critical for performance. This advance knowledge allows the system to pre-configure power management strategies, allocating power efficiently before workloads execute, thereby supporting larger cluster sizes within the same power budget.
3Device complexity
If existing methods are used to manage network power, then device complexity is maintained, but power optimization effectiveness is insufficient
Solution Approach 1:
The patent introduces a power management intermediary layer that sits between the application layer and the physical network infrastructure. This intermediary analyzes application communication patterns and translates them into power management decisions for network links, providing effective power optimization without requiring complex changes to either the application code or the underlying network hardware.
Solution Approach 2:
The patent implements feedback mechanisms where network link performance and power consumption are continuously monitored, and this information is fed back into the power management system. The system uses this feedback to adjust power allocation dynamically, improving optimization effectiveness while maintaining relatively simple system architecture through automated closed-loop control.
Data Source
AI summary
Examples of the present disclosure include methods, devices, and/or systems. Identifying network communication patterns can include analyzing a distributed computer program of a network, estimating virtual network communication traffic based on the analysis, and mapping the virtual network communication traffic to a physical network link. Identifying network communications patterns can also include identifying the network communication pattern and categorizing the physical communication network link based on an estimated communication intensity of the mapped communication traffic and the network communication pattern. Identifying network communication patterns can further include optimizing an energy used by the network based on the categorization.