RDMA
RDMA is a direct memory access technology that allows a computer to transfer data directly to the memory of another computer over a network without passing through the CPU or operating system of either party. RDMA has the characteristics of low latency, high bandwidth, and low CPU load. Using RDMA can effectively reduce the latency of data transmission and improve data transmission efficiency. It is very suitable for application environments such as high-performance computing or data centers to improve the overall performance of the network.
RDMA is particularly important in networks that require fast, massive parallelism, such as high-performance computing, autonomous driving, etc. It can analyze and process large amounts of data simultaneously and transmit them with low latency and high transmission rates. This makes autonomous driving possible because, during autonomous driving, the system needs to analyze and respond to complex road conditions. In addition, RDMA can also be used for computing between nodes in a cluster and for latency-sensitive database workloads.
How Does RDMA Work
RDMA is the implementation of the transmission protocol through a RoCE-enabled network card. Before the transmission, a connection needs to be established between the two RDMA nodes, which needs to be implemented through RoCE. When using RDMA, the host application puts its data packets into a memory buffer and then mobilizes the network or storage driver. The remote direct memory access stack allows the hardware adapter to access the raw buffer directly. This allows it to bypass many traditional software stacks and avoid large memory buffer copy operations.
Such steps can effectively reduce latency and speed up data transmission. When both devices use RDMA for transmission, the data transmission speed will be faster than when only one device supports RDMA.
The Characteristics of RDMA
RDMA has some significant features that support it in improving the overall performance of the network and improving the efficiency of data transmission.
Low latency: Since RDMA allows direct data transmission between remote memory and local memory, bypassing steps such as the operating system and CPU, this significantly reduces communication latency.
High bandwidth: Since data is read and written directly in memory, RDMA can make better use of network bandwidth, increase bandwidth utilization efficiency, and increase the speed of data transmission.
Reduce CPU load: Since RDMA data transmission does not need to be processed by the CPU, it directly reads data from the computer's memory and writes it to the memory of another computer, reducing CPU intervention and freeing up processor resources.
Zero copy: RDMA can directly allow data transmission between buffers without involving the network software stack, avoiding data replication between network layers.
Advantages of RDMA
RDMA has several advantages:
Kernel bypass: RDMA allows data transfer directly in user space without the need for context switching through the kernel. This approach effectively reduces system call overhead and improves the overall efficiency of data transfer.
Message-based transactions: RDMA manages data in the form of discrete messages rather than continuous streams. This approach simplifies the task of applications when processing streaming data, making it more efficient and simpler to split data streams into different messages or transactions.
Scatter/gather entry support: RDMA has natural support for multiple scatter/gather entries. It can read data from multiple memory buffers and combine them into a stream, or distribute a data stream to multiple memory buffers. This flexibility improves the efficiency and flexibility of data processing.
Conclusion
RDMA technology improves network performance by reducing network latency and increasing network bandwidth. It is indispensable for high-tech. If you have any questions about RDMA, please feel free to please feel free to contact QSFPTEK's CCIE/HCIE engineers at [email protected]. Our engineers will provide you with comprehensive support.