“We use the identical bodily structure as anybody with prime of rack after which leaf and backbone swap,” Dudy Cohen, vice chairman of product advertising at DriveNets, advised Community World. “However what occurs between our prime of rack, which is the swap that connects NICs (community interface playing cards) into the servers and the remainder of the community isn’t based mostly on Clos Ethernet structure, fairly on a really particular cell-based protocol. [It’s] the identical protocol, by the best way, that’s used within the backplane of the chassis.”
Cohen defined that any information packet that comes into an ingress swap from the NIC is lower into evenly sized cells, sprayed throughout the whole material after which reassembled on the opposite aspect. This method distinguishes DriveNets from different options that may require specialised parts resembling Nvidia BlueField DPUs (information processing models) on the endpoints.
“The material hyperlinks between the highest of rack and the backbone are completely load balanced,” he mentioned. “We don’t use any hashing mechanism… and that is why we are able to include all of the congestion avoidance throughout the material and don’t want any exterior help.”
Multi-site implementation for distributed GPU clusters
The multi-site functionality permits organizations to beat energy constraints in a single information middle by spreading GPU clusters throughout areas.
This isn’t designed as a backup or failover mechanism. Lasser-Raab emphasised that it’s a single cluster in two areas which can be as much as 80 kilometers aside, which permits for connection to totally different energy grids.
The bodily implementation sometimes makes use of high-bandwidth connections between websites. Cohen defined that there’s both darkish fiber or some DWDM (Dense Wavelength Division Multiplexing) fibre optic connectivity between the websites. Sometimes the connections are bundles of 4 800 gigabit ethernet, appearing as a single 3.2 terabit per second connection.