In-Network Multicast Switching to Reduce Core Thread Overhead
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multi-processor systems face inefficiencies in broadcast and multicast communications due to the need for direct point-to-point messaging, which becomes burdensome as system scale increases, lacking configurability for distributed memory architectures with multiple network hops, and requiring significant thread resources.
Innovation Solution
Implementing a network of configurable switches with ISA extensions and hardware support for interrupt generation, using collective virtual circuits and per-port register descriptions to efficiently propagate multicast or broadcast operations across cores, optimizing the use of thread resources and reducing network hops.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If direct point-to-point messaging is used for broadcast and multicast operations, then communication can be implemented in existing multi-processor systems, but thread resource usage increases significantly and performance deteriorates as system scale increases
Solution Approach 1:
The patent introduces a network switch as an intermediary device that handles broadcast and multicast message distribution. Instead of having the source processor directly manage point-to-point messages to multiple destinations, the network switch receives the message from the source processor and automatically distributes it to all intended recipients using its internal forwarding table. This mediator approach transfers the distribution burden from processor threads to the network switch, significantly reducing thread resource usage and improving communication performance in scaled systems.
2Adaptability or versatility
If software-level implementation of broadcast and multicast is used, then flexibility and configurability are maintained, but thread resources are consumed in the foreground and overhead increases
Solution Approach 1:
The network switch is configured with forwarding tables that enable it to autonomously determine message distribution paths without requiring ongoing software intervention. Once the forwarding table is set up, the network switch independently handles all broadcast and multicast distribution decisions, eliminating the need for foreground thread resources to manage each message transmission. This self-service capability reduces thread resource consumption while maintaining configurability through the initial table setup.
3Ease of operation
If network architecture supports broadcast and multicast with group addresses, then message distribution to device groups is enabled, but granularity of configurability is insufficient for distributed memory architectures with multiple network hops
Solution Approach 1:
The patent implements a hierarchical forwarding table structure where different levels of the network hierarchy can have customized forwarding rules. Each network switch in the distributed memory architecture maintains its own forwarding table with entries tailored to its specific position and connectivity in the network. This allows fine-grained configurability at each network hop, enabling precise control over message distribution paths through multiple switches while maintaining ease of operation through standardized switch interfaces.
Data Source
AI summary
Examples described herein relate to a first group of core nodes to couple with a group of switch nodes and a second group of core nodes to couple with the group of switch nodes, wherein: a core node of the first or second group of core nodes includes circuitry to execute one or more message passing instructions that indicate a configuration of a network to transmit data toward two or more endpoint core nodes and a switch node of the group of switch nodes includes circuitry to execute one or more message passing instructions that indicate the configuration to transmit data toward the two or more endpoint core nodes.