Validating AI Data Center Front-End Networks
视频
As large language models (LLMs) continue to scale, the complexity of the supporting infrastructure increases, presenting challenges beyond GPU performance alone. A key concern is validating the data pipelines that feed these models, particularly at the front end of AI data centers. This includes securely ingesting data from distributed cloud and customer environments and enabling fast, low-latency transfers between virtual machines and hosts within the data center. The session highlights the critical importance of front-end networks in delivering data efficiently to GPUs during LLM training. Two major traffic patterns are discussed: north-south traffic, which handles data ingestion from external sources and requires strong security and low latency; and east-west traffic, which focuses on internal VM communication and demands ultra-low latency and high throughput. To address these challenges, Keysight introduces CyPerf, a software-based traffic generator that emulates real-world application traffic and measures key performance indicators such as latency, bandwidth, connections per second, and security. CyPerf plays a crucial role in validating whether network infrastructure can meet the rigorous demands of AI workloads by simulating application-layer behavior, not just network throughput. The presentation emphasizes the importance of testing all seven layers of the OSI model, particularly the application layer, to ensure systems are production-ready and not reliant on user feedback for critical performance issues.