SmartNIC Remote Memory Offloading for Passive Memory Nodes

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing user-space runtime library based remote memory systems introduce significant CPU computing and management overhead at memory nodes due to the need for management units and network protocol stack CPU polling, limiting complete disaggregation of memory and computing resources.

Innovation Solution

A SmartNIC offloading based remote memory system that utilizes a System-on-a-Chip (SoC) on the SmartNIC to offload management of remote memory nodes, enabling in-network management and one-sided RDMA access, reducing CPU involvement at memory nodes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If user-space runtime library based remote memory access is implemented, then application semantics and fine-granularity access are enabled, but CPU resources at memory nodes are heavily consumed due to management unit overhead and network protocol stack polling

Engineering Contradiction:
Improveapplication semantics utilizationVSAvoidCPU resource consumption
Core Design Contradiction:
Ease of operationVSUse of energy by moving object

Solution Approach 1:

The patent extracts the management unit and network protocol stack polling functions from the memory node CPU, relocating them to the SmartNIC. This allows the memory node to operate passively without active CPU intervention, while the SmartNIC handles all management operations including address translation and RDMA packet processing

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The SmartNIC acts as an intermediary between the compute node and memory node, handling all remote memory access management. It translates object IDs to physical addresses, manages RDMA queue pairs, and processes network packets, freeing the memory node CPU from these overhead tasks

Inventive Principle:
Principle #24Intermediary (Mediator)

2Manufacturing precision

If management units are deployed at remote memory nodes to process remote access requests, then object-granularity access is achieved, but computing overhead is introduced at memory nodes

Engineering Contradiction:
Improveaccess granularityVSAvoidmemory node structure
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The management unit is extracted from the memory node and relocated to the SmartNIC. The SmartNIC now performs object ID to physical address translation and manages fine-granularity access control, while the memory node structure is simplified to only contain passive memory arrays without complex management logic

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The SmartNIC serves as an intermediary that handles all address translation and access control operations. It maintains the object ID to physical address mapping tables and performs precise address translation for fine-granularity access, eliminating the need for complex management units at memory nodes

Inventive Principle:
Principle #24Intermediary (Mediator)

3Reliability

If CPU polling is used for network packet reception at memory nodes, then reliable data reception is ensured, but CPU resources are always occupied

Engineering Contradiction:
Improvedata reception reliabilityVSAvoidCPU availability
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The network protocol stack and packet reception polling functions are extracted from the memory node CPU and relocated to the SmartNIC. The SmartNIC autonomously handles RDMA packet reception, parsing, and processing, ensuring reliable data reception without occupying memory node CPU resources

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The SmartNIC performs self-service by autonomously handling all network communication tasks including packet reception, address translation, and memory access operations. It manages its own RDMA queue pairs and completion queues without requiring CPU intervention, freeing the memory node CPU for other productive tasks

Inventive Principle:
Principle #25Self-service

Data Source

PatentUS20250220078A1Smartnic offloading based remote memory system
Publication Date: 2025.07.03 SHANGHAI JIAOTONG UNIV
  • US20250220078A1 patent drawing
  • US20250220078A1 patent drawing
  • US20250220078A1 patent drawing

AI summary

A SmartNIC offloading based remote memory system pertains to the field of communication technology and includes a compute node, a SmartNIC and a memory node. The compute node sends a remote memory access request to the SmartNIC. The SmartNIC is used to parse the remote memory access request and access the memory node. The memory node initializes and saves remote memory. The SmartNIC includes a remote memory access request handling module, a data address management module and an RDMA communication front end, and the memory node includes an RDMA communication back end. Use of the remote memory system includes an initialization stage, an application registration stage and a running stage. The system offloads management overhead from the memory node to the SmartNIC, achieving in-network management of remote memory.