Direct Interconnect Fabric for Data Center Switching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data center architectures are inefficient due to the use of multiple levels of Ethernet or LAN switches, which are expensive and slow down message and packet passage, and require different chassis form factors for compute nodes and switches, leading to increased complexity and cost.
Innovation Solution
Implementing a direct interconnect fabric that virtualizes peripherals such as Network Interface cards, Ethernet cards, hard disks, BIOS, and consoles, allowing processors to share these devices and eliminating the need for rack and aggregation switches by performing switching functions within the fabric, thereby flattening the hierarchy and reducing the number of LAN switches.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If multiple levels of Ethernet or LAN switches are used to connect compute nodes and storage nodes, then network connectivity and device access are achieved, but packet and message delays increase and system cost increases
Solution Approach 1:
The patent extracts the switching function from external Ethernet/LAN switches and relocates it into the interconnect fabric within each chassis. This eliminates the need for separate rack switches and aggregation switches, directly reducing packet delays and the number of network devices while maintaining full connectivity between compute and storage nodes.
Solution Approach 2:
The patent merges the external LAN switching function with the internal interconnect fabric. By combining these previously separate functions into a unified fabric-based switching mechanism, the system eliminates multiple switching layers and reduces both packet delays and the total number of switches required.
2Adaptability or versatility
If different chassis form factors are used for compute nodes and switches, then functional separation is achieved, but data center complexity and cost increase
Solution Approach 1:
The patent makes the interconnect fabric universal by enabling it to perform both internal node communication and external switching functions. This multi-functionality allows a single chassis form factor to accommodate compute nodes, storage nodes, and switching capabilities, eliminating the need for separate specialized chassis for each function.
Solution Approach 2:
The patent segments the switching function from the physical chassis structure. Instead of requiring dedicated switch chassis, the switching capability is segmented as a virtual function within the fabric, allowing any chassis to perform switching operations based on software configuration rather than hardware specialization.
3Ease of operation
If rack switches and aggregation switches are deployed in hierarchical levels, then network management and traffic routing are achieved, but packet delays and system cost increase
Solution Approach 1:
The interconnect fabric acts as an intermediary that eliminates the need for intermediate rack switches and aggregation switches. By providing direct fabric-based routing between any compute or storage node, the system removes intermediate switching hops while maintaining centralized management capabilities through software-defined networking.
Solution Approach 2:
The patent transitions from a hierarchical switching architecture (multiple vertical levels) to a flattened fabric-based architecture (horizontal peer-to-peer connectivity). This dimensional change allows any node to communicate directly with any other node through the fabric, eliminating the need for traffic to traverse multiple hierarchical levels and reducing packet delays.
Data Source
AI summary
A data center has several dis-aggregated data clusters that connect to the Internet through a firewall and load-balancer. Each dis-aggregated data cluster has several dis-aggregated compute/switch/disk chassis that are connected together by a mesh of Ethernet links. Each dis-aggregated compute/switch/disk chassis has many processing nodes, disk nodes, and I/O nodes on node cards that are inserted into the chassis. These node cards are connected together by a direct interconnect fabric. Using the direct interconnect fabric, remote I/O and disk nodes appear to the operating system to be located on the local processor's own peripheral bus. A virtual Ethernet controller and a virtual generic peripheral act as virtual endpoints for the local processor's peripheral bus. I/O and disk node peripherals are virtualized by hardware without software drivers. Rack and aggregation Ethernet switches are eliminated using the direct interconnect fabric, which provides a flatter, dis-aggregated hierarchy.