Multiple Request Notification Network for Coherent Mesh Interconnects
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current coherent mesh interconnect networks are limited in managing global ordering of memory requests, as they only allow for single request notifications per node per notification window, failing to efficiently handle multiple requests from a requester node.
Innovation Solution
A multiple request notification network is integrated with a notification network and a multiple request count network to enable the notification of multiple requests from a requester node and at least one request from other nodes per notification window, using switch-point circuitry with control logic to manage and broadcast request indicators and counts across interconnected nodes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single request notification per node per notification window is used, then the notification network complexity is reduced, but the productivity of memory request processing is limited
Solution Approach 1:
The notification network is segmented into multiple independent notification windows (first notification window, second notification window, etc.). Each window can independently notify multiple requests from the same node, allowing parallel processing of multiple requests without increasing the complexity of a single notification window. This segmentation enables high throughput while maintaining manageable network complexity.
Solution Approach 2:
The system uses periodic notification windows to broadcast requests in discrete time intervals. Each notification window operates independently and periodically, allowing multiple requests from the same node to be notified across different windows. This periodic structure provides a systematic way to handle multiple requests without requiring complex continuous arbitration mechanisms.
2Productivity
If multiple requests from a single node are notified in parallel, then the productivity of memory access is improved, but the difficulty of detecting and measuring request ordering increases
Solution Approach 1:
Request indicators are assigned sequential identifiers (first request indicator, second request indicator, etc.) before being broadcast in parallel across different notification windows. This preliminary ordering establishes a global sequence that all nodes can independently track, making it easy to detect and measure request ordering even when requests are processed in parallel. The sequential labeling provides a clear measurement framework.
Solution Approach 2:
The system provides feedback to all nodes about which node has priority to issue requests in each notification window. This feedback mechanism allows nodes to adjust their request issuance timing based on observed system state, enabling efficient parallel processing while maintaining detectable global ordering through the sequential request indicators.
3Reliability
If a priority node selection mechanism is implemented, then the global ordering of requests is improved, but the device complexity increases
Solution Approach 1:
Priority node selection is performed periodically at the beginning of each notification window rather than continuously. The control logic selects which node has priority to issue requests for the upcoming window, providing consistent global ordering without requiring complex continuous arbitration. This periodic approach simplifies the control logic while maintaining reliability.
Solution Approach 2:
The system determines the priority node in advance at the beginning of each notification window before requests are issued. This preliminary determination of priority establishes a clear ordering framework that all nodes can follow, ensuring global ordering consistency without requiring complex real-time decision-making logic during request processing.
Data Source
AI summary
A data processing system includes a plurality of processing unit. Each processing unit includes notification storage circuitry configured to store a notification indicator corresponding to each processing unit which indicates whether the processing unit has an outstanding coherent memory request, and multiple request storage circuitry configured to store a multiple request indicator corresponding to each processing unit which indicates whether the processing unit has more than one outstanding request. The data processing system also includes an interconnect network coupled between the processing units and configured to broadcast coherent memory requests from a requesting processing unit of the plurality of processing units to other processing units of the plurality of processing units.