Vivek Bagmar

Zero-Loss by Design: A Field-Tested Calibration Methodology for Lossless AI/GPU Ethernet Fabrics

Abstract:

Large AI and GPU clusters challenge conventional Ethernet assumptions because synchronized collective operations generate intense microbursts, incast, many-to-many traffic, and highly latency-sensitive congestion. In this environment, a network may report zero packet drops yet still deliver suboptimal GPU performance due to deep queues, excessive PFC propagation, or unstable tail latency. This talk presents a field-tested methodology for engineering lossless AI/GPU Ethernet as a calibrated operating envelope rather than a simple configuration checkbox. The approach treats losslessness as the result of carefully tuning the interaction among traffic classes, NIC behavior, buffers, ECN, PFC, queue thresholds, and workload characteristics. Calibration begins with a clean baseline covering MTU, priority mapping, QoS trust boundaries, ECN capability, and PFC alignment. Network load is then progressively increased to observe when queues build, ECN marking begins, PFC activates, and packet loss occurs. Synthetic stress profiles are used to reproduce realistic AI traffic patterns, including incast, synchronized many-to-many traffic, elephant flows, bursts, and mixed- workload contention. The methodology emphasizes early ECN as the primary congestion-control mechanism, with PFC reserved as an emergency safeguard. It also helps identify pause storms, threshold mismatches, silent microburst loss, asymmetric congestion, head-of-line blocking, and misleading throughput-test results. Finally, network telemetry is correlated with workload events, collective latency, job consistency, and tail behavior to establish a repeatable calibration framework that improves both network efficiency and GPU workload performance.

Brief Profile:

Vivek Bagmar is a Proof of Concept Engineer at Arista Networks in the San Francisco Bay Area, with more than 12 years of experience in data center and enterprise networking. In his current role, he designs and validates large-scale, high-performance network architectures, helping customers evaluate advanced solutions across EVPN-VXLAN, AI-driven networking, and cloud-scale data center fabrics. Vivek brings extensive experience working in complex, multi-vendor environments and previously held senior technical and sales engineering roles at Juniper Networks. He is recognized for his hands-on leadership in proof-of-concept engineering, network automation, and telemetry-driven performance optimization. A dual JNCIE-certified expert in Enterprise and Data Center networking, Vivek combines deep technical expertise with the ability to translate complex network architectures into clear business value. He is passionate about advancing modern networking technologies and helping organizations build scalable, resilient, and future-ready infrastructure.