Distributed Packet Switching in Source Routed Cluster Servers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Cluster server networks with centralized switches face issues of fault susceptibility, latency, bandwidth bottlenecks, and scalability limitations due to large routing tables and tree topology constructions, which are costly and inefficient.
Innovation Solution
Implementing a distributed packet switching system with a mesh network topology using deterministic routing and source-route-control MAC addresses, along with local caching and a master translation table for link layer address translations, to enable transparent virtualization and reduce storage requirements at each node.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a centralized L2 switch is used for inter-node communications, then packet switching is simplified, but reliability deteriorates due to single point of failure and performance deteriorates due to latency and bandwidth bottlenecks
Solution Approach 1:
The patent segments the centralized switch into distributed switches at each compute node. Each node maintains local routing tables and forwarding logic, eliminating the single centralized switching component and its associated failure point while distributing the switching function across multiple nodes.
Solution Approach 2:
The patent transitions from a two-dimensional tree topology to a three-dimensional mesh topology, adding horizontal interconnects between nodes. This dimensional expansion provides multiple routing paths and eliminates the root switch bottleneck, improving both reliability and performance.
2Device complexity
If a tree topology is used for L2 Ethernet switch networks, then network structure is simplified, but performance deteriorates due to root switch bandwidth limitation requiring large expensive root switches
Solution Approach 1:
The patent transforms the hierarchical tree topology into a flat mesh topology with three-dimensional connectivity. This eliminates the single root switch bottleneck by distributing switching capacity across all nodes, allowing any node to forward traffic to any destination through multiple paths.
Solution Approach 2:
In the mesh topology, every compute node serves multiple functions: it acts as both an endpoint and a switching node. Each node can forward traffic for other nodes, making the network fully distributed and eliminating the need for a specialized high-capacity root switch.
3Adaptability or versatility
If each switch stores routing information for every path, then routing flexibility is improved, but device complexity increases due to large routing tables limiting scalability
Solution Approach 1:
Instead of storing complete routing information for all possible paths at every switch, the patent implements partial routing tables at distributed nodes. Each node maintains routing information relevant to its local context and uses source routing to determine paths, reducing the routing table size at each individual node while maintaining overall network routing flexibility.
Data Source
AI summary
A cluster compute server includes nodes coupled in a network topology via a fabric that source routes packets based on location identifiers assigned to the nodes, the location identifiers representing the locations in the network topology. Host interfaces at the nodes may be associated with link layer addresses that do not reflect the location identifier associated with the nodes. The nodes therefore implement locally cached link layer address translations that map link layer addresses to corresponding location identifiers in the network topology. In response to originating a packet directed to one of these host interfaces, the node accesses the local translation cache to obtain a link layer address translation for a destination link layer address of the packet. When a node experiences a cache miss, the node queries a management node to obtain the specified link layer address translation from a master translation table maintained by the management node.