Layered Full-Mesh Node Interconnects for Scalable Memory Pooling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computing systems face challenges in efficiently managing increased memory requirements due to the growing number of users and complex applications, necessitating improved network management for scalable and efficient resource pooling.
Innovation Solution
A system with a switch baseboard that interconnects nodes via a first layer for full-mesh connections within clusters and a second layer for inter-cluster connections, supporting cache-coherent protocols, and utilizing switches with different bandwidth ports to form hyper-torus connections, enabling efficient data transfer and memory pooling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If nodes are interconnected using traditional network topologies, then the system can support basic memory requirements, but the system cannot efficiently handle increased memory requirements from growing users and complex applications
Solution Approach 1:
The network is divided into multiple layers: a first layer providing full-mesh connections within clusters, and a second layer providing inter-cluster connections. This segmentation allows the system to scale memory capacity by adding clusters while maintaining efficient access patterns through the hierarchical structure.
Solution Approach 2:
The patent introduces a two-layer network topology that adds a dimensional aspect to node interconnections. The first layer handles intra-cluster traffic while the second layer handles inter-cluster traffic, creating a multi-dimensional communication path that improves efficiency as the system scales.
2Ease of operation
If a full-mesh connection is implemented across all nodes, then any node can directly access any other node, but the network complexity and connection requirements become unmanageable as the system scales
Solution Approach 1:
The full-mesh connectivity is segmented into cluster-level full-mesh connections (first layer) and inter-cluster connections (second layer). This reduces overall complexity by limiting full-mesh requirements to smaller clusters while using a more manageable connection pattern between clusters.
Solution Approach 2:
The network topology transitions from a single-plane full-mesh to a two-layer hierarchical structure. The first layer provides direct full-mesh access within clusters, while the second layer provides structured inter-cluster access, reducing complexity while maintaining accessibility.
3Speed
If high bandwidth connections are provided for all node pairs, then data transfer speed is maximized, but the cost and complexity of providing sufficient bandwidth across the entire network becomes prohibitive
Solution Approach 1:
Different bandwidth characteristics are assigned to different network layers. The first layer (intra-cluster) and second layer (inter-cluster) can have different bandwidth capacities, allowing optimization of bandwidth resources based on local traffic patterns and requirements rather than uniformly provisioning high bandwidth everywhere.
Solution Approach 2:
The bandwidth provisioning is separated into two dimensional layers, allowing independent optimization of bandwidth resources for intra-cluster and inter-cluster communications. This enables efficient resource allocation where high bandwidth is provided where most needed while reducing bandwidth requirements in other areas.
4Adaptability or versatility
If the network is designed to support a fixed number of nodes, then the network configuration is simple, but the system cannot scale to support growing memory requirements
Solution Approach 1:
The network is organized into modular clusters with standardized internal full-mesh connectivity. New clusters can be added to the second layer without reconfiguring existing clusters, enabling scalable growth while maintaining simple, repeatable configuration patterns within each cluster.
Solution Approach 2:
The first layer full-mesh connections are pre-configured within each cluster before deployment. This preliminary configuration allows clusters to be independently assembled and then integrated into the larger network through the second layer, simplifying the scaling process.
Data Source
AI summary
Provided are systems and methods for network management. A system includes a switch baseboard including switches to interconnect nodes via a first layer connecting a first cluster of the nodes into full-mesh-connected nodes, a second cluster of the nodes into full-mesh-connected nodes, a third cluster of the nodes into full-mesh-connected nodes, and a fourth cluster of the nodes into full-mesh-connected nodes, and a second layer including inter-cluster connections to connect a first node of the first cluster to a first node of the second cluster by one hop, and a first node of the third cluster by one hop.