Subnet Number Aliasing for InfiniBand Scalability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing InfiniBand network architecture faces limitations in scalability and flexibility due to the 16-bit Local Identifier (LID) space, which restricts the number of unicast addresses, leading to challenges in routing and load-balancing across multiple subnets, especially in large-scale cloud computing environments.
Innovation Solution
The implementation of subnet number aliasing through the assignment of Fabric Local Subnet Numbers (FLSNs) and the use of a global fabric manager to dynamically reconfigure the network, allowing for alias FLSN assignment and monitoring, thereby enhancing scalability and flexibility in routing and load-balancing across multiple subnets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If hierarchical mapping structure is used to map 128 bit DGID values to 16 bit DLID values, then scalability in terms of total number of subnets and end nodes is improved, but the ability to route individual destinations independently is severely limited
Solution Approach 1:
The patent segments the addressing space into two parts: a subnet identifier portion and a local identifier portion. The subnet identifier identifies the subnet, while the local identifier identifies individual destinations within that subnet. This segmentation allows hierarchical routing at the subnet level while maintaining independent routing capability at the local level, resolving the contradiction between scalability and routing flexibility.
Solution Approach 2:
The patent introduces an intermediary mapping mechanism where subnet identifiers are mapped to ranges of local identifiers. This intermediary layer enables efficient hierarchical lookup while preserving the ability to route to specific destinations within subnets, balancing scalability requirements with independent routing capabilities.
2Device complexity
If 16 bit LID space is used for unicast addresses, then addressing within a single subnet is simplified, but the number of available addresses is limited to 48K
Solution Approach 1:
The patent extends the addressing space from a single 16-bit dimension to a two-dimensional structure: subnet identifier (portion of DGID) and local identifier (16-bit LID). This dimensional expansion allows the system to address a much larger number of destinations by combining subnet-level identification with local-level identification, overcoming the 48K address limit while maintaining simple 16-bit local addressing within each subnet.
3Measurement precision
If content addressable memory based lookup hardware is used for mapping DGID to DLID, then mapping accuracy is improved, but hardware scalability and cost are severely constrained
Solution Approach 1:
The patent segments the lookup process into two stages: first, lookup the subnet identifier in a routing table to determine the appropriate subnet; second, use the local identifier portion for final destination identification within that subnet. This segmentation replaces the need for a single large content addressable memory with smaller, more scalable lookup tables, maintaining mapping accuracy while improving hardware scalability.
Solution Approach 2:
The patent extracts the subnet identifier portion from the full 128-bit DGID address and uses it as the key for routing table lookup. This extraction allows the system to perform efficient subnet-level routing without requiring content addressable memory to handle the entire 128-bit address, significantly reducing hardware complexity and cost while maintaining accurate destination mapping.
Data Source
AI summary
System and method for supporting subnet number aliasing in a high performance computing environment. In accordance with an embodiment, a fabric member can be assigned, by a global fabric manager, an alias fabric local subnet number in order to keep a fabric running after a fabric reconfiguration. The alias fabric local subnet number can be assigned for a period of time, the period of time being static, configurable, or indefinite.