Artificial Intelligence Generated Content (AIGC) technology leverages advanced generative models to produce a wide range of content, including natural language text, images, and audio. Within the architecture of AIGC networks, the Network Interface Card (NIC) serves a vital function as the primary device that facilitates connectivity between computers and networks. The primary responsibilities of the NIC include the efficient transmission of data generated by the computer to the network, as well as the receipt of incoming data. NIC devices are essential in guaranteeing the high performance and reliability of AIGC networks, thereby providing a robust foundational support system for data transmission and network connectivity.
Reasons for the dual uplink of the network card
Dual NIC bonding is a network architecture that establishes simultaneous connections between two physical network interface cards (NICs) of a server or network device and distinct upper-level devices or switches.
In traditional single network card architecture, interruptions in AIGC training tasks can occur due to failures in either fibre optic connections or switches. Such interruptions can result in increased training costs and adversely affect customer branding. Furthermore, during switch upgrades, it is necessary to migrate AIGC training operations in advance, which can pose significant challenges to user experience, system stability, and network operation and maintenance. Conversely, the dual uplink architecture for network cards enhances reliability by connecting the two ports of all network cards on the server to separate switches. By binding these two ports to form a bonded port, this architecture ensures seamless service provision. Consequently, in the event of a failure of one uplink link or its corresponding access layer switch, traffic can be rerouted to the alternate port, thereby preventing disruptions to training tasks.
The dual upper-link architecture design mitigates the risk of a single point of failure associated with connecting the network card to a singular switch, thereby significantly enhancing the robustness of the overall system interconnection. This design also facilitates the hot upgrade of switches within cluster systems, thereby improving the convenience of network operation and maintenance as well as the processes of functional iterations
Network Card Dual Uplink Architecture Network Solution
The following outlines several dual uplink architecture solutions for network interface cards that are currently supported by the existing switch infrastructure:
● Dual IP Network Card
In this configuration, each port of the network interface card is assigned two distinct IP addresses, effectively dispersing traffic through different pathways according to the network card's configuration. This setup allows the network card to be virtually represented as two separate network cards, thereby leveraging the mature IP forwarding capabilities of switches. In the event of a failure of one port or IP address, the alternative port or IP address remains operational. The dual IP configuration for network cards is a versatile and efficient networking solution, applicable across various settings. However, it is important to note that certain collective communication libraries provide inadequate support for dual IP configurations, potentially resulting in performance degradation, particularly with multiple Queue Pairs (QPs). Furthermore, this solution necessitates the allocation of double the number of IP addresses for each network interface card, which could lead to inefficient use of IP address resources.
● Stack
The de-stacking solution is an innovative approach introduced by our organization. This solution involves the integration of network interface cards and switches to establish an aggregation port. On the network interface card side, ARP/ND messages are broadcast simultaneously by both ports, enabling the two connected switches to learn the network card's ARP/ND information concurrently. The connected switches subsequently convert these ARP/ND entries into Border Gateway Protocol (BGP) routes, which are then disseminated to other devices.
The stacking approach allows the method of business access to remain unchanged, eliminating the necessity for a physical connection between the two switches while effectively accommodating dual network card uplink access.
● Stacking + Dual Planes
The stacking combined with dual planes solution builds upon the stacking concept by partitioning the switch into distinct forwarding planes. Each dual uplink port of the network interface card is allocated to different network planes, meaning that the two ports connect to separate switches, each linked to different planes.
By employing the de-stacking and dual-plane configuration, the network card's sending end can ensure even traffic distribution across the two ports, leading to a balanced flow of network traffic at the receiving side's access layer switch. This notably diminishes the likelihood of hash polarization.
Furthermore, this design of dual uplink and dual-plane access effectively increases the maximum expansion scale of a single cluster within a two-layer CLOS network, yielding advantages such as simplified overall cluster communication topology, reduced latency, and lower operational costs.
see more information : https://www.ruijienetworks.com/support/tech-gallery/multi-dimensional-comparison-and-analysis-of-aigc-network-card-dual-uplink-technical-architecture