Cache Coherence Conflict Resolution in HPI Interconnects
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current interconnect architectures in high-performance computing systems face challenges in efficiently managing communication between multiple processors and devices, leading to performance bottlenecks and increased power consumption, especially as the number of processing units and devices grows.
Innovation Solution
The development of a High Performance Interconnect (HPI) system with a layered protocol architecture, including a coherence protocol, routing layer, link layer, and physical layer, which enables efficient data transfer and cache coherence management across multiple processors, using point-to-point links and virtual channels to optimize communication and reduce power usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If traditional multi-drop buses are used for interconnect communication, then device compatibility is maintained, but communication performance and bandwidth capacity deteriorate
Solution Approach 1:
The interconnect architecture is segmented into multiple independent point-to-point links rather than a single shared bus. Each link has dedicated bandwidth and can operate independently, eliminating the bottleneck of shared medium access. This segmentation allows parallel communication paths that scale with the number of processors without increasing overall system complexity.
Solution Approach 2:
The patent transitions from a one-dimensional shared bus architecture to a multi-dimensional hierarchical interconnect structure with multiple levels (e.g., first level interconnect for fast communication, second level for coordination). This dimensional expansion provides additional communication pathways and reduces contention for resources.
2Productivity
If the number of processing units and devices is increased to boost computing power, then processing capability improves, but interconnect communication burden and power consumption increase
Solution Approach 1:
The interconnect protocol enables continuous data transfer without idle waiting periods. By maintaining persistent connections and using flow control mechanisms, the system keeps communication channels actively utilized rather than idle, reducing the energy wasted by repeatedly establishing and dismantling connections for each transaction.
Solution Approach 2:
The system performs preliminary actions by establishing connection states and buffering mechanisms in advance. Flow control credits and buffer allocations are prepared beforehand, allowing data to be transmitted continuously without frequent pauses for handshaking, thereby reducing the energy consumption associated with repeated protocol overhead.
3Reliability
If more data is transferred between processors to achieve coherence, then data consistency is maintained, but communication overhead and latency increase
Solution Approach 1:
The coherence protocol segments data transfers into smaller units that can be processed in parallel. By dividing large data transfers into multiple smaller transactions that can occur simultaneously across different links, the system reduces total latency while maintaining coherence. Multiple processors can perform coherence operations independently without waiting for a single sequential broadcast.
Data Source
AI summary
A coherence protocol message is sent corresponding to a particular cache line. A potential conflict involving the particular cache line is identified and a forward request is sent to a home agent to identify the potential conflict. A forward response can be received in response to the forward request from the home agent and a response to the conflict can be determined.