DPU Workload Migration via Utilization Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In computing device ecosystems, resources on one node are often over-utilized while other nodes have under-utilized resources, making it challenging to identify and migrate workloads effectively to balance DPU utilization.
Innovation Solution
A method for DPU workload management involves obtaining DPU utilization values, identifying workloads with high CPU utilization or strong correlation with node DPU utilization as migration candidates, and migrating these workloads to nodes with available resources to balance DPU utilization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If workloads are migrated based on traditional resource utilization metrics, then CPU or memory balancing may be improved, but DPU utilization balancing remains ineffective
Solution Approach 1:
The patent implements a feedback mechanism by continuously monitoring DPU utilization metrics and using this information to dynamically adjust workload migration decisions. The system measures actual DPU utilization on source nodes and uses this feedback to identify appropriate migration targets, thereby resolving the contradiction between general resource balancing and specific DPU utilization balancing.
Solution Approach 2:
The patent replaces traditional CPU/memory-based workload migration mechanisms with a DPU-utilization-aware migration system. Instead of relying on conventional resource utilization metrics, the system substitutes DPU-specific utilization measurements as the primary driver for migration decisions, enabling precise control over DPU resource distribution across the cluster.
2Productivity
If DPU utilization monitoring is implemented at the node level, then overall DPU balancing improves, but identification of specific migration candidates becomes complex
Solution Approach 1:
The patent changes the parameter used for workload identification from traditional CPU utilization to a composite metric that incorporates DPU utilization correlation. By transforming the selection criterion to prioritize workloads with strong DPU utilization patterns, the system simplifies candidate identification while maintaining effective DPU balancing.
Solution Approach 2:
The patent introduces an intermediary layer that analyzes the relationship between workloads and DPU utilization. This intermediary component calculates correlation metrics between workload performance and DPU resource consumption, thereby simplifying the identification of migration candidates that will most effectively balance DPU utilization across nodes.
3Productivity
If workloads are migrated to balance DPU utilization, then DPU resource distribution improves, but system complexity increases
Solution Approach 1:
The patent implements a universal workload migration framework that can handle multiple objectives simultaneously. The system is designed to balance DPU utilization while also considering CPU, memory, and other resource metrics through a unified migration decision-making process, thereby managing system complexity through multi-functionality rather than separate specialized systems.
Data Source
AI summary
Techniques described herein relate to systems and methods for data processing unit (DPU) workload management. Such methods may include obtaining, by a DPU workload manager, a DPU utilization value for a first node in a device ecosystem; making a first determination, by the DPU workload manager and using the DPU utilization value, that DPU utilization of the first node is above a DPU utilization threshold configured for the first node; identifying, in response to the first determination, a workload executing on the first node as a migration candidate based at least in part on a central processing unit (CPU) utilization value associated with the workload; initiating a migration of the migration candidate to a second node in the device ecosystem; and obtaining, after the migration completes, a second DPU utilization value for the first node to determine whether the second DPU utilization value is below the DPU utilization threshold.