Multi-DPU Seamless Offload via Capability-Based Bonding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current load balancing techniques using data processing units (DPUs) in computer networks face limitations in throughput and processing capacity due to reliance on single DPU cards, which are constrained by silicon availability and architecture, leading to suboptimal performance and inability to meet demands for higher throughput applications.
Innovation Solution
A system and method for seamless offloading of workloads to multiple DPUs based on capabilities within a threshold, using a library-level approach that selects and bonds multiple DPU cards to surpass individual DPU limitations, leveraging APIs and support libraries to determine and utilize DPUs with matching or complementary capabilities for load balancing, thereby optimizing workload distribution and performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If single DPU card is used for load balancing, then device complexity is reduced, but throughput and processing capacity are limited by silicon availability and architecture
Solution Approach 1:
The patent combines multiple DPU cards into a unified load balancing system, merging their processing capabilities to achieve throughput beyond what a single DPU can provide. The system bonds multiple DPUs together while presenting a single interface to applications, effectively combining their silicon resources and architectural capabilities.
Solution Approach 2:
The patent segments the load balancing workload across multiple DPU cards, dividing the processing tasks among different hardware units. This segmentation allows the system to handle higher throughput by distributing work across multiple silicon instances rather than overloading a single DPU.
2Productivity
If multiple DPU cards are bonded to surpass limitations, then throughput and processing capacity improve, but system complexity increases at library level
Solution Approach 1:
The patent introduces a library-level intermediary that manages the complexity of bonding multiple DPUs. This intermediary layer handles DPU selection, capability matching, and workload distribution, shielding applications from the underlying system complexity while enabling multiple DPUs to work together seamlessly.
Solution Approach 2:
The system dynamically changes parameters such as DPU selection criteria and capability thresholds based on workload requirements. By adjusting these parameters at the library level, the system can optimize processing capacity without requiring hard-coded complexity in the application layer.
3Productivity
If DPUs are selected based on capability matching within threshold, then workload distribution is optimized, but selection and configuration complexity increases
Solution Approach 1:
The patent implements self-service mechanisms where the library automatically performs DPU capability assessment, selection, and configuration based on predefined thresholds and workload requirements. This automation eliminates manual selection complexity while optimizing workload distribution across suitable DPUs.
Solution Approach 2:
The system incorporates feedback loops that monitor DPU performance and capability utilization, using this information to dynamically adjust workload distribution and DPU selection. This feedback mechanism optimizes productivity while keeping selection logic automated and threshold-based rather than manually complex.
Data Source
AI summary
Systems and methods herein are for seamless offload of a workload to data processing units (DPUs), where one or more processing unit receive a selection of a first one of the plurality of DPUs to perform the workload, and perform a background operation to select second ones of the plurality of DPUs based, at least in part, on capabilities associated with the first one of the plurality of DPUs being within a threshold, where the workload is to be performed in a load balancing arrangement of the first one and second ones of the plurality of DPUs.