NVIDIA® Mellanox® Multi-Host technology enables next generation Cloud, Web 2.0 and high-performance data centers to design and build new scale-out heterogeneous compute and storage racks with direct connectivity between multiple hosts and the centralized network controller. This enables direct data access with the lowest latency to significantly improve densities and maximizes data transfer rates.
NVIDIA’s Multi-Host technology provides the ability to allow multiple compute or storage hosts to connect into a single interconnect adapter, by separating the adapter PCIe bus into several independent interfaces. Each interface connects to a separate host CPU—with no performance degradation and improves operational agility and efficiency.
Image courtesy of Facebook: Yosemite V3 multi-host server
The ConnectX-6 Dx SmartNIC adds significant improvements to NVIDIA Multi-Host applications by offering advanced QoS features that ensure complete isolation among the multiple hosts connected to the NIC, and by achieving superior fairness among the hosts. ConnectX-6 Dx is equipped with a large shared ingress buffer designed to hold bursty traffic when the incoming traffic rate exceeds the rate of traffic that can be absorbed by the host. The shared buffer configuration allows a single flow to consume the entire buffer space when no competition for buffer usage exists. When multiple flows consume the shared ingress buffer, the buffer is allocated in a manner that ensures that each host and each priority within the hosts are served in a fair manner.
Multi-Host technology allows for different CPU architectures (x86, Power, ARM, etc.) and each host to be kept completely independent from the others. Hosts sharing the same network interface save on switch ports, cables, real estate and power.
Multi-Host technology allows direct connectivity from multiple hosts to the network, enabling direct data access which lowers latency and overcomes performance bottlenecks to maximize data processing and transfer rates.
We're here to help you build the most efficient, high performance network.
