Dell Enterprise SONiC network operating system
Dell Technologies has been an innovator in the open-source and disaggregation arena for many years, starting with the integration of running third-party networking operating systems on Dell branded networking equipment and, more recently, running open-source code such as SONiC.
Dell Technologies has created a world-class, enterprise-grade version of the open-source SONiC stack. The Dell Enterprise SONiC stack provides a full suite of enterprise, cloud, edge, and campus features.
This stack delivers scalable, high-performance, multi-tenancy connections for Hyper-Converged Infrastructure (HCI), Converged Infrastructure (CI), and AI workloads.
Dell Enterprise SONiC, starting with version 4.2.1 on the Z9664F-ON, has implemented a series of networking features that facilitate the deployment of a single to multi-rack GPU cluster.
The following figure shows the initial Dell Enterprise SONiC for AI networking feature set. Each feature aims to address performance, quality-of-service, and ease of management for an AI deployment.
To address the high-performance requirements of an AI workload, Dell networking provides a switch product portfolio with 400GbE and 800GbE switching capacity.
With RoCEv2, chaotic and demanding AI network traffic patterns can be guaranteed in a congested fabric. AI traffic can be classified and marked as a high strict priority traffic in the fabric. The classification as a strict priority ensures that AI traffic is assigned into a specific output queue that is serviced before any other traffic in the fabric. This first-priority classification ensures smooth data switching from end-to-end, in this case from GPU-to-GPU.
With cut-through switching, latency sensitive traffic such as AI is switched within the network device without waiting to receive the entire AI data packet before switching the packet to its destination.
And finally, load balancing features such as enhanced hashing and dynamic load balancing address points of congestion to uniformly distribute AI traffic across multiple 400GbE ports as needed.
These Dell Enterprise SONiC features create a solid foundation on which a lossless, high-performance, and scalable Ethernet fabric can be created.