InfiniBand Switching Tag Placement for Reduced Latency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high cost and complexity of InfiniBand network switches due to large address ranges and memory requirements make it economically challenging for entry-level business users to deploy and manage InfiniBand networks, particularly for small-scale networks with limited nodes, leading to increased latency and packet processing overhead.
Innovation Solution
A network manager detects nodes using explorer procedures and selects address field sizes for data packets based on the number of nodes, configuring switches with a switching tag at the packet's start to minimize switching complexity and reduce addressing requirements, thereby optimizing switching resources and performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If InfiniBand network switches use large address ranges (16-bit LIDs) to support unique addressing, then addressing capability is improved, but device complexity and memory requirements increase substantially
Solution Approach 1:
The patent segments the addressing function into two parts: a compressed switching tag (5-8 bits) for routing decisions and a full 16-bit LID for unique identification. The switching tag is inserted at the start of packets for quick switch lookup, while the original LID is preserved in the packet payload for destination identification. This segmentation allows switches to use smaller, simpler address tables while maintaining full addressing capability end-to-end.
Solution Approach 2:
The patent introduces an intermediary component (the subnet manager and compression algorithm) that translates between the compact switching tag format and the full 16-bit LID format. The subnet manager dynamically compresses LIDs to switching tags based on the actual number of active nodes in the network, creating an intermediate representation that reduces switching complexity while preserving addressing capability.
2Adaptability or versatility
If InfiniBand network switches support large address ranges, then addressing coverage is improved, but memory requirements and cost increase
Solution Approach 1:
The patent segments the address handling function so that switches only need to store and process compact switching tags (5-8 bits) in their forwarding tables, rather than full 16-bit LIDs. This reduces the memory required in switch address tables by a factor of 2-4 times while maintaining the ability to address all nodes in the network through the compressed tag space.
Solution Approach 2:
The patent changes the parameter of address field size from fixed 16-bit LIDs to variable-length switching tags (5-8 bits) that are dynamically determined based on the number of active nodes in the network. The subnet manager calculates the appropriate tag size as ceil(log2(N)) where N is the number of active nodes, optimizing memory usage while maintaining sufficient addressing coverage.
3Ease of manufacture
If InfiniBand networks are deployed for small-scale networks with limited nodes, then deployment cost is reduced, but latency increases due to unnecessary addressing overhead
Solution Approach 1:
The patent applies partial action by using only the necessary portion of the addressing capability. Instead of always using full 16-bit LIDs, the system uses only ceil(log2(N)) bits as switching tags where N is the actual number of active nodes. For small networks, this means using only 5-8 bits out of 16, reducing packet processing overhead and latency while maintaining unique addressing for all active nodes.
Solution Approach 2:
The patent dynamically changes the addressing parameter based on network size. The subnet manager determines the optimal switching tag size as ceil(log2(N)) where N is the number of active nodes, and configures all switches to use this optimized tag size. This parameter adaptation reduces the addressing overhead in small networks from 16 bits to 5-8 bits, decreasing packet processing time and latency.
Data Source
AI summary
A network manager, configuring for detecting network nodes and configuring network switches, determines addressing field lengths to be used for addressing the network nodes and switching data packets between the network nodes based on the number of detected network nodes. The network manager detects the network nodes by exploring the network according to prescribed explorer procedures. The network manager selects a size of address fields to be used for switching data packets traversing the network, based on the number of detected network nodes. The network manager configures each network switch within the network to switch the data packets based on a switching tag having the selected size and positioned at the start of the packet. Hence, each network switch is able to generate forwarding decisions based on the switching tag at the beginning of each received data packet. The switching tag is distinct from, and substantially smaller than, the existing destination address field. Hence, switching complexity can be minimized for relatively small networks having minimal addressing requirements, reducing latency and simplifying forwarding decisions within the network switches.