Welcome to the end to end life cycle presentation of an AI optimized fabric using Cisco Nexus dashboard. Today we trace the path from provisioning and lossless settings to real time checks, showing how intelligent packets frame controls.00:05Open flows in your AI backend.00:20This session is prepared by Aporva Takar, a customer success specialist, and aims to give you detailed insights to improve your network efficiency. Today's agenda provides a structured overview of our session in great detail. 1st, we'll cover AI and ML infrastructure, then discuss Cisco Nexus 9300.00:22Hardware solutions followed by exploring various network architecture options and AI fabric building blocks. We will address traffic load balancing challenges and then move on to discuss dynamic load balancing and intelligent packet flow.00:42Finally, we will clearly demonstrate these concepts in action. Designing for AI requires evolving beyond traditional data center architecture. We use a two tier approach. The front end network handles standard user to app interactions. While the back end network, the very core of the AI cluster is a.00:56High bandwidth, lossless fabric for low latency energy GPU communication essential for intensive training. Segregating these systems prevents management traffic from bottlenecking AI tasks.01:16Distributed training is offline and compute intensive requiring high node to node bandwidth to maximize training speed. Conversely, production inference operates live. Prioritizing high availability and low latency for real time responses. Differentiating these design options is essential for infrastructure planning.01:27Cisco's Nexus 9300 series 400 G switches deliver very high speed throughput and deep buffering, which are critical for lossless AI fabrics. The 800 G switch provides 51.2 tb of capacity and 256 mb of buffers. Effectively doubling previous.01:47Capacities and ensuring future proofing with ultra ethernet readiness. Nexus Dashboard integrates existing ACI or VXLAN EVPN fabrics. We discussed two design options, the classic design in which nodes connect to a single leaf and the rail optimized design in which GPU Nick pairs connect to unique rails across multiple.02:07Believes to optimize all reduced patterns. Both employ high performance non blocking BGP fabrics. To achieve a lossless fabric, we employ row CV two to bypass the CPU and we use PFC and ECN to actively manage buffer occupancy with DCQC and RED providing robust congestion control.02:27Advanced load balancing ensures optimal, highly efficient link utilization.02:47Next, we focus on load balancing techniques. Traditional ECMP five tuple hashing often fails in AI networks today, producing traffic polarization in which a single spine becomes overburdened. Dynamic load balancing DLB addresses this by monitoring link utilization in real time and rerouting traffic using flowlet hashing.02:51Real world data show that DLB maintains around 50 % utilization across ports worldwide maximizing throughput. Effective traffic flow management is pivotal in optimizing an AI network as demonstrated in the current slide for this deployment.03:12On one side, ECMP static pinning provides predictability by assigning flows to fixed paths, aiding and troubleshooting and stable behavior in traditional data centers for IT teams and operators. In contrast for AI workloads with long lived flows, static pinning can cause congestion. This approach truly favors purpose.03:30Packet spraying. Instead of forcing a flow to a single path, packets are spread across multiple available spine links.03:50By using all available paths packet spring mitigates bottlenecks by pooling the total bandwidth of all spines. The result is that static pinning offers the predictability required for standard traffic. Packet spring reduces congestion for AI data flows.03:57Cisco Nexus switches use highly intelligent packet techniques, offering advanced load balancing strategies tailored to diverse workloads ensuring both robust consistency and high performance in any AI fabric. Combining Cisco per packet DLB with nvidia spectrum X adaptive routing yields a 37 % performance gain.04:13Improving jane's fairness index to 1.00. With Cisco per packet load balancing and NVIDIA spectrum X, the results are shown here. This is what we mean by optimized traffic distribution for AI networks. By distributing packets across all links, we eliminate congestion and ensure there are NO unkno.04:33To utilize paths. For AI and ML workloads, such uniformity leads to consistent high speed training maximizing infrastructure ROI and efficiency.04:53Supported hardware includes the specific nexus 9300 platforms and nvidia's Bluefield three super Nix. Firmware must fully support the LLDP TLV handshake mechanism for adaptive routing capabilities. Licensing is per switch. Use XF two or XF three license family based on your hardware.05:03For PID. This diagram illustrates a classic leaf spine fabric with both 100 gbps and 400 gbps connections. The diagram shows high speed 400 gbps, backbone links in dashed green, and 100 gbps links in dashed light blue. The solid blue lines clearly delineate the.05:22Active traffic paths originating from the compute hosts, specifically hosts one through five.05:42Communicating with a centralized storage node at interface E1/eleven on leaf two oh six. This many to one arrangement can lead to network congestion. These screenshots show the real time traffic rates on the links between leaf two oh six and the various spine switches. Here we can see that the links are not used evenly in practice. Some are overused while others are underused.05:47Monitoring ECN and output discards indicates congestion management. We observe several output discards arising from the leaf toward the storage end host along the path. Moreover, the packets are not marked with ECN even though there is congestion and packet drops. The storage host trans.06:10Of its PPP pause frames to notify about the network congestion.06:27In the upcoming steps, we will examine how cisco's DLB and other capabilities of lossless Ethernet are able to address and resolve this issue. Now, at this point, please move over to the Nexus dashboard to configure dynamic load balancing with priority based flow control. To provision, navigate to manage fabrics in the Nexus dashboard, choose create.06:32New land fabric, then select fabric type as AI and choose AI routed subtype.06:52Name the fabric to find BGP ASN 400 G fabric and telemetry settings. In the advanced settings stage, we have the option to customize further, but for this demonstration, we'll stick with the recommended defaults. These defaults give a balance of flexibility and reliability.07:02Designed with industry best practices. We will click next in order to advance to the summary before beginning the process of fabric creation as part of the standard workflow. We will click next in order to advance to the summary before beginning the process of fabric creation as part of the standard workflow.07:19On the following page, you will be able to review the accuracy of all the information presented. After you have verified the summary, submit to trigger the automated creation process. With the submission now complete, we are presently entering the fabric creation phase. At this point the fabric has been created.07:39Once the fabric has been created, begin by locating all switches by using seed IP addresses and the credentials as a next step. From site one, select L1L2S1 and S two, and then add these devices to the fabric, assign spine roles to all of the added devices.07:58S one and S two acknowledge the instruction to recalculate and deploy in order to validate the topology and to ensure correctness across the connected devices. Open the actions menu, then choose recalculate and deploy. Carefully reviewing the options before proceeding.08:20Once we calculate and deploy are started, the nexus dashboard begins recalculating the required configurations for the switches we have selected. This ongoing process ensures every configuration item aligns exactly with the intended fabric design, thereby supporting an AI optimized infrastructure.08:37On this page, we can preview configurations, compare the running state, the pending state, and the expected state, and then proceed to deploy. In this screen, the Nexus dashboard provides a detailed view of the pending configuration for site 1L1 device.08:53We are currently examining the site 1L1 switch in particular at this time. The dashboard presents the exact commands that will be implemented, including quality of service, queuing policies, and MTU settings tailored for our AI network. After the preview, we have now proceeded with deploying the configurations across all devices.09:08Once deployment is complete.09:29The switch show commands tool can be used to verify operational status directly from the Nexus dashboard. We have run the show run ipqos command to verify the QoS settings on our site's L one switch. Live output confirms that our QoS policies, including dross EV two classifications, are correctly set.09:31Next, on host connectivity, we manage interfaces via the connectivity tab. We have selected Ethernet one 5th on site 1L1, which is connected to our storage host in this laboratory, and we are editing interface E15 for AI workloads to enable PFC and cost configuration.09:50Again, we can observe here that Nexus Dashboard is generating the necessary configurations for the Ethernet port one 5th. This display shows the current state of the switch port configuration. Here we explore the pending configuration view for ongoing AI workloads here now today. This display shows clearly the Nexus dashboard's command.10:10Needs to move from the current configuration to your desired design. Let's move ahead to the expected config view. This projection shows in detail how the configuration will appear once all pending changes are deployed for the interface.10:30We are presently positioned in deployment, where the Nexus dashboard is actively applying the optimized settings to the Ethernet interface one 5th at site 1L1. This automation smoothly migrates from the expected config to the actual switch, removing manual entry and loading.10:47Oring human errors. Once the config is applied, production shows balanced traffic across interfaces delivering production grade performance for AIML applications. Leaf interfaces now use every interface equally. Thanks for watching. Subscribe for more advanced networking topics.11:04
