RDMA Network Card Message Forwarding for Low-Latency Overlays
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The mismatch between network bandwidth, processor speed, and memory bandwidth in data center networks leads to significant latency and resource overhead, exacerbated by the need for data copying and transmission through multiple software layers, which traditional TCP/IP technology cannot efficiently address.
Innovation Solution
A message forwarding method and apparatus based on Remote Direct Memory Access (RDMA) that utilizes a combination of hardware and software modules in a network card to encapsulate and forward messages, reducing latency and resource overhead by enabling direct memory access without operating system intervention, applicable in cloud network overlay architectures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional TCP/IP technology is used for data packet processing, then the operating system and software layers can handle network traffic, but this occupies large amounts of server resources and memory bus bandwidths, and imposes a heavy burden on the CPU and memory of servers
Solution Approach 1:
The patent extracts the network packet processing function from the operating system software layer and implements it directly in hardware (network card). The network card independently completes tasks such as packet encapsulation, decapsulation, forwarding, and filtering without involving the CPU and memory system, thereby resolving the contradiction between reliable network traffic handling and server resource occupation.
Solution Approach 2:
The network card acts as an intermediary between the physical network and the operating system. It performs all necessary network protocol processing (TCP/IP stack functions) in hardware, serving as a mediator that eliminates the need for software-layer intervention and reduces the burden on server resources while maintaining full network functionality.
2Reliability
If data is transmitted through multiple software layers with copy-move operations between system memory, processor caches, and network controller caches, then network communication can be achieved, but this causes significant latency and network delay effects
Solution Approach 1:
The patent extracts data transmission operations from the multi-layer software processing path and implements them directly in hardware. The network card performs encapsulation, forwarding, and other operations without requiring data to be copied between system memory, processor caches, and network controller caches, thereby eliminating the time loss associated with multiple memory access operations while maintaining full network communication capability.
Solution Approach 2:
The patent replaces the mechanical process of data copying and moving between multiple memory layers (software-based) with a direct hardware-based transmission path. The network card's hardware architecture enables data to be processed and transmitted without undergoing multiple copy operations, substituting the inefficient mechanical copying process with a streamlined hardware operation that significantly reduces latency.
3Productivity
If hardware modules are used for message encapsulation and forwarding, then bandwidth and processor overhead are reduced, but the flexibility of message processing may be compromised
Solution Approach 1:
The network card's hardware module is designed with multi-functionality to handle various message processing tasks including encapsulation, decapsulation, forwarding, and filtering of different packet types. This universal design enables the hardware to adapt to diverse networking requirements while maintaining high bandwidth efficiency, resolving the contradiction between productivity and adaptability.
Data Source
AI summary
A message forwarding method and apparatus based on Remote Direct Memory Access, and a network card and a device. The method includes: acquiring payload data corresponding to an RDMA message to be forwarded; encapsulating a message header of the payload data by using a hardware module in a network card node to obtain a processed message; and forwarding the processed message by using the hardware module or a software module in the network card node.