Hybrid Network Switch for Multi-Node Cache Coherency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Multiprocessor computer systems face performance limitations due to the requirement that access rights of processors change in the order of received snoops in broadcast protocols, leading to increased latency, especially in systems with out-of-order execution, prefetching, multiple cores, and multi-threading. Additionally, directory-based protocols suffer from high latency and network congestion, and the complexity and cost of maintaining directories.
Innovation Solution
A hybrid network switch that dynamically determines the transmission mode for coherence requests based on a mode table, allowing either broadcast or point-to-point mode, to optimize network utilization and reduce latency by broadcasting only when necessary and using point-to-point connections for widely shared data or read-only regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If broadcast protocol is used to maintain cache coherency, then all processors can observe coherence requests in the same order, but access rights must change in the order of received snoops which increases latency
Solution Approach 1:
The patent applies dynamics by enabling the system to switch between broadcast mode and point-to-point mode based on the coherence request type. The mode table dynamically determines the transmission mode, allowing the system to adapt to different scenarios and optimize performance while maintaining correctness.
Solution Approach 2:
The patent segments coherence requests into different types (exclusive access, shared access, read-only) and applies different transmission modes to each segment. This segmentation allows the system to avoid unnecessary broadcasts for read-only requests while maintaining broadcast for exclusive access scenarios, thereby reducing latency without compromising coherency.
2Reliability
If broadcast protocol is used for coherence requests, then cache coherency is maintained, but network bandwidth is wasted for read-only data
Solution Approach 1:
The patent applies partial action by broadcasting coherence requests only when necessary (for exclusive access or shared access scenarios) while using point-to-point mode for read-only requests. This partial broadcast approach avoids wasting network bandwidth on unnecessary transmissions while still maintaining cache coherency where needed.
3Loss of energy
If directory-based protocol is used to reduce broadcast overhead, then network congestion is reduced, but latency increases and system complexity increases
Solution Approach 1:
The patent applies universality by creating a hybrid system that supports both broadcast and point-to-point modes within a single coherence protocol framework. The mode table serves as a universal controller that directs coherence requests to the appropriate transmission mode, combining the advantages of both broadcast (low latency for exclusive access) and point-to-point (bandwidth efficiency for read-only) protocols without implementing separate directory-based systems.
4Device complexity
If broadcast protocol is used for all coherence requests, then implementation is simple, but performance degrades for out-of-order execution systems
Solution Approach 1:
The patent applies parameter changes by modifying the transmission mode parameter based on the coherence request type. The mode table stores parameters that determine whether broadcast or point-to-point mode should be used, allowing the system to optimize performance for out-of-order execution while maintaining relatively simple implementation through a structured approach to mode selection.
Data Source
AI summary
A node in a multi-node system includes a memory, an active device that includes a cache, an interface that sends and receives coherency messages on an inter-node network coupling the node to another node, and an address network that communicates address packets between the devices in the node. In response to receiving a coherency message from the other node requesting an access right to a coherency unit, the interface sends an address packet on the address network. The address packet is a first type of address packet if the coherency unit is in the modified global access state in the node and a second type of address packet otherwise. If the active device is the owner of the coherency unit, the active device responds to the first type of address packet and ignores the second type of address packet.