Switch Fabric Control Logic for Parallel Broadcast and Collect Operations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Fabric primitives in multi-core, 10 Gigabit Ethernet clusters, particularly broadcast and collect operations, are inefficient when performed serially, leading to high latency and processing burdens, and have been neglected in new fabric technologies despite their critical role in parallel processing.
Innovation Solution
Implementing a switch fabric with control logic that aggregates responses from multiple destination host devices into a single aggregated response, using data plane and control plane components to identify, count, and discard responses, and generate aggregated packets, thereby optimizing collect operations and reducing latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If fabric primitives are performed serially, then processing is simpler, but latency increases and productivity decreases
Solution Approach 1:
The patent segments the fabric primitive processing into distinct phases: broadcast phase where the source host sends requests to multiple destination hosts, and collect phase where responses are gathered and aggregated. This segmentation allows parallel execution of multiple operations simultaneously, reducing overall latency while maintaining manageable processing complexity through structured phase separation.
Solution Approach 2:
The patent merges multiple response packets from different destination hosts into a single aggregated response packet. The switch fabric combines these responses in parallel during the collect phase, reducing the number of packets that need to be processed individually and thereby decreasing latency without significantly increasing processing complexity.
2Productivity
If fabric primitives are performed serially, then device complexity is lower, but productivity decreases
Solution Approach 1:
The switch fabric is designed with universal control logic that can handle both broadcast and collect operations through the same infrastructure. The same switch fabric ports and control mechanisms used for data transmission also handle response aggregation, eliminating the need for dedicated hardware for each operation type and achieving parallel processing throughput without proportionally increasing device complexity.
Solution Approach 2:
The source host preliminarily establishes the collect operation parameters and destination host list before initiating the broadcast. The switch fabric is pre-configured with the distribution tree for efficient response gathering. This preliminary setup enables high-speed parallel collection of responses without requiring complex real-time decision-making during the actual data gathering phase.
3Loss of time
If responses from multiple hosts are collected individually, then data integrity is maintained, but latency increases
Solution Approach 1:
The switch fabric implements feedback mechanisms to track the collection status of responses from multiple destination hosts. The control logic monitors which hosts have responded and ensures all expected responses are gathered before completing the collect operation. This feedback system maintains data integrity by verifying complete response collection while reducing latency through efficient parallel gathering and aggregation of responses.
Data Source
AI summary
Techniques are described for optimizing broadcast and collect primitives in switch fabrics. A switch is described for use in a switch fabric that includes a plurality of switches configured to interconnect a plurality of host devices. The switch includes a plurality of ports configured to receive and transmit data, and switch logic configured to facilitate transfer of the data among the ports. The switch logic is configured to implement a collect function in which responses from multiple destination host devices are collected and aggregated for transmission to a source host device. A first portion of the switch logic is configured to identify, count, and discard the responses in a data plane of the switch. A second portion of the switch logic is configured to generate an aggregated response in a control plane of the switch and introduce the aggregated response into the data plane for transmission toward the source host device.