On-Demand Path Update for MSDC Scaling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In massively scaled data center architectures, maintaining globally-unique indexes for all end ports in leaf switches becomes impractical due to limited forwarding table sizes, leading to inefficient traffic forwarding and compromised Quality of Service (QoS).
Innovation Solution
Implementing a multi-layer addressing technique that uses domain indexes to direct traffic to egress domains, with a transition from less-specific paths defined by domain indexes to more-specific paths defined by fabric system ports, allowing for dynamic allocation and reuse of local indices for efficient traffic routing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If ingress leaf switches maintain globally-unique index information for all end ports, then traffic forwarding can be performed using static index lookups, but the forwarding table size requirements become impractical and scale poorly
Solution Approach 1:
The patent segments the globally-unique index into two components: a domain index (identifying the egress domain) and a local index (identifying the specific egress port within that domain). This segmentation allows ingress leaf switches to store only domain indexes in their forwarding tables, dramatically reducing table size requirements while maintaining efficient lookup performance. The local index is obtained dynamically through path learn requests when needed.
Solution Approach 2:
The patent introduces a hierarchical dimension to the addressing scheme by organizing indexes into multiple levels: domain level (global domain index) and port level (local index within domain). This multi-dimensional approach replaces the traditional flat globally-unique index structure, enabling scalable forwarding in large data centers without requiring prohibitively large forwarding tables at each switch.
2Adaptability or versatility
If a static globally-unique index is used for all end ports, then the indexing scheme is simple to implement, but it does not scale to massively scaled data center architectures
Solution Approach 1:
The indexing scheme is segmented into domain indexes and local indexes. Domain indexes provide global uniqueness across the data center, while local indexes provide port identification within domains. This segmentation enables the system to scale to massively large architectures by distributing the indexing burden across multiple levels rather than requiring each switch to maintain complete global index information.
Solution Approach 2:
The patent introduces dynamic path learn requests that allow the system to adapt to changing network conditions and topology. When traffic needs to be forwarded to a specific egress port, the ingress switch dynamically obtains the local index through a path learn request mechanism, rather than relying on static pre-configured global index information. This dynamic approach enables scalability while maintaining forwarding efficiency.
Data Source
AI summary
In one embodiment, a copy of an original packet of a traffic flow is created at an ingress leaf node of a cloud switch. The ingress leaf node forwards the original packet along a less-specific path through the cloud switch, the less-specific path based on a domain index of an egress domain for the original packet. The copy of the original packet is modified to create a more specific path learn request packet. The ingress leaf node forwards the more specific path learn request packet along the less-specific path through the cloud switch. The ingress leaf node received back a more specific path learn request reply packet that includes an indication of a fabric system port. The ingress leaf node then programs a forwarding table based on the indication of the fabric system port, to have subsequent packets of the traffic flow forwarded along a more-specific path.