What Is Lossless Ethernet?_
Lossless Ethernet extends standard Ethernet with mechanisms like Priority Flow Control (PFC) to eliminate packet loss due to congestion, which is critical for storage and high-performance computing traffic. It relies on a lossless fabric where switches and NICs cooperate to pause and resume traffic flows without dropping frames. This is distinct from standard Ethernet's best-effort delivery, as it ensures data integrity for sensitive protocols like RDMA over Converged Ethernet (RoCE).
Technical Details
Lossless Ethernet typically implements PFC per traffic class, creating separate virtual lanes for different data types to avoid head-of-line blocking. It requires end-to-end configuration of pause thresholds and buffer allocations across all switches and adapters to maintain a lossless state. The fabric must be carefully tuned to prevent deadlock or unfairness, often using quantized congestion notification (QCN) or similar algorithms. Standards like IEEE 802.1Qbb define PFC, but actual deployment parameters follow OEM specifications based on the specific hardware and workload.
How Leviathan Systems Works with Lossless Ethernet
In our field work, Lossless Ethernet is configured on the scale-out network connecting GPU nodes via Ethernet fabrics (e.g., RoCEv2), where we set PFC parameters on switches and NICs per the OEM guide. It is a common requirement for RoCEv2 deployments in H100 and Blackwell racks, ensuring RDMA traffic between GPUs across the data center does not suffer retransmissions.
Related service
Network Testing & Commissioning →