RDMA: A New Technology to Solve High-latency Problems in High-performance Computing
When dealing with applications with high IO concurrency and low latency, such as high-performance computing and big data analysis, the traditional TCP/IP software and hardware architecture is gradually unable to meet today's needs. Why? The reason is that traditional TCP/IP network communication relies on the kernel to transfer data. This communication method has high data movement and data copy overhead.
The emerged RDMA (Remote Direct Memory Access) technology has significant advantages. Its technical principle is to allow data transmission between applications without kernel intervention, and to access data in memory directly through the network interface, effectively solving the problem of high latency caused by the kernel during data transmission. Because of this, RDMA is especially suitable for scenarios that have extremely high performance requirements and are extremely sensitive to latency.
RDMA Model VS. Conventional Model
The Two Heroes of RDMA Network: InfinibandTM and RoCE
Currently, there are two main types of RDMA networks, namely InfinibandTM and RoCE (RDMA over Converged Ethernet).
InfinibandTM is a network designed specifically for RDMA, which guarantees reliable transmission for hardware, provides higher bandwidth and lower latency. It is technologically advanced, but expensive, and requires network cards, switches, optical modules and DACs that support IB.
Remote Direct Memory Access (RDMA) is the capacity for remote memory management, which permits the direct movement of data between the application memories of servers, without the intervention of any CPU. RDMA over Converged Ethernet (RoCE) is a means to provide this efficient data transfer with very low delays on lossless Ethernet networks. You can use ordinary switches to build a network, but the network card must support RoCE.
Types of RDMA Network
Zhaolong QSFP-DD 800G DAC that Supports RDMA Network
As one of Zhaolong’s LONGSPEEDTM Ethernet high-speed interconnection solutions, QSFP-DD 800G DAC has demonstrated excellent performance in supporting RDMA network, becoming a powerful weapon to reduce latency. Compared with other similar products, it has unique advantages. Its advanced design and technical implementation can ensure high-speed and stable data transmission in the RDMA network environment, and escort high-performance computing with minimal latency. No matter applied in InfinibandTM or RoCE network environment, Zhaolong’s QSFP-DD 800GDAC can give full play to its own characteristics, provide a low-latency experience far beyond the norm, and help solve the stringent requirements for low latency in high-performance computing.
Zhaolong’s QSFP-DD800 800 Gbps to QSFP-DD800 800 Gbps DAC
The advantages and features of Zhaolong QSFP-DD800 (800G) DAC products are:
Using Zhaolong's independently developed LONGSPEEDTM Twin-Ax HSC series high-speed bare wire, the cable performance is excellent;
The transmission encoding adopts PAM4-112G technology, supports no resonance (Suck-out) before 40GHz, and is backward compatible with QSFP-DD400, QSFP56, and QSFP28;
The product complies with the QSFP-DD800 MSA standard and the IEEE 802.3ck specification as well. The maximum transmission speed for Zhaolong QSFP-DD800 (800G) DAC product is 800Gbps, and it is backward compatible with 400G products.
Passive design, low power consumption. Insertion loss is less than 19.75dB@26.56GHz, and COM (Channel Operating Margin) is better than 4dB under the condition of superimposing 8 near-end crosstalk and 7 far-end crosstalk.
Supports 26-30AWG, 0-70℃application, and supports 0.5-2.0m short-distance to long-distance applications
Provide LSZH products and customized solutions (such as LSZH outer sheath, pull-tabs’label color, etc.)
English
简体中文
