Smart Endpoint Protocol Acceleration for Disaggregated Data Centers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computing systems face inefficiencies in resource utilization and management when using hyperconverged servers, as different types of resources (processors, storage, accelerators) have different refresh rates and usage patterns, leading to suboptimal performance and increased costs.
Innovation Solution
The implementation of a data center architecture with disaggregated resources, where resources such as compute, memory, and storage are allocated dynamically to form managed nodes, allowing for independent upgrading and optimization of each resource type, and utilizing a smart endpoint and offload complex for flexible protocol acceleration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If resources are aggregated in hyperconverged servers, then resource sharing is simplified, but resource utilization efficiency deteriorates due to different refresh rates and usage patterns
Solution Approach 1:
The patent segments resources by function and refresh rate into separate physical pools (compute resources, storage resources, accelerator resources) rather than aggregating them in hyperconverged servers. Each resource type can be independently managed and allocated based on its specific refresh rate and usage patterns, resolving the contradiction between simplified sharing and utilization efficiency.
Solution Approach 2:
The patent implements dynamic resource allocation where the resource allocation manager can flexibly assign different types of resources to different virtual machines based on real-time workload requirements. This dynamic allocation allows compute resources to be shared among multiple VMs while storage and accelerator resources are allocated based on their specific characteristics, optimizing overall utilization.
2Adaptability or versatility
If resources are disaggregated into separate pools, then independent upgrading and optimization is enabled, but system complexity increases
Solution Approach 1:
The patent creates a universal resource allocation framework that can manage multiple types of disaggregated resources (compute, storage, accelerators) through a common interface and management mechanism. The resource allocation manager provides multi-functional capabilities to handle different resource types uniformly, reducing the perceived complexity despite physical disaggregation.
Solution Approach 2:
The resource allocation manager acts as an intermediary layer between the disaggregated resource pools and the virtual machines. This intermediary abstracts the complexity of managing multiple separate resource pools, presenting a unified view to users while enabling independent upgrading and optimization of each resource type underneath.
3Device complexity
If fixed function endpoints are used in PCIe devices, then hardware configuration is simplified, but protocol flexibility and acceleration capability deteriorates
Solution Approach 1:
The patent replaces fixed function PCIe endpoints with dynamic smart endpoints that can be programmatically configured to support multiple protocols. The smart endpoint includes a protocol parser and accelerator that can be programmed via firmware to handle different protocols (NVLink, CXL, PCIe, etc.), providing flexibility while maintaining hardware efficiency through dedicated acceleration capabilities.
Solution Approach 2:
The patent enables protocol flexibility by changing the operational parameters of the PCIe endpoint through programmable firmware. The smart endpoint can dynamically adjust its behavior and protocol handling based on firmware configuration, allowing the same hardware to adapt to different protocols and acceleration requirements without physical reconfiguration.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Technologies for flexible I/O protocol acceleration include a computing device (1602) having a root complex (1624), a smart endpoint (1632) coupled to the root complex (1624), and an offload complex (1634) coupled to the smart endpoint (1632). The smart endpoint (1632) receives an I/O transaction that originates from the root complex (1624) and parses the I/O transaction based on an I/O protocol and identifies an I/O command. The smart endpoint (1632) may parse the I/O transaction based on endpoint firmware that may be programmed by the computing device (1602). The smart endpoint (1632) accelerates the I/O command and provides a smart context to the offload complex (1634). The smart endpoint (1632) may copy the I/O command to memory of the smart endpoint (1632) or the offload complex (1634). The smart endpoint (1632) may identify protocol data based on the I/O command and copy the protocol data to the memory of the smart endpoint (1632) or the offload complex (1634). Other embodiments are described and claimed.