Google adds IPv6 only AI clusters to GKE in preview
Each node gets 65,536 IPv6 addresses for pods, and Google tells that speed was the same as IPv4 in its internal tests.
Google added single stack IPv6, a new type of network address, to Google Kubernetes Engine (GKE), its system that controls groups of computers. Nodes, the computers, and pods, the units of software on them, use only IPv6 and no private IPv4 addresses. It is in preview for the Cloud TPU, a Google AI processor, and for CPU node pools.
How it works
Google tells that large AI systems use all private IPv4 addresses before they use all computers or power. Pods connect to IPv4 only endpoints through DNS64 and NAT64, which change one address type to the other. Each node gets 65,536 pod addresses.
The numbers
In internal tests, Google tells that pod start, latency and data speed were the same as with IPv4. It tells that load balancing and DNS had no cost that it could measure. Google gives no figures.
What is not known
Only the Cloud TPU and some CPU types work at this time, and Google plans more processors. Usual GKE limits for pods on a node apply. No test by a different group is given.
Insight: Removing the address limit does not remove the next limit
In November, Google told that one cluster has a limit of 65,536 network identities, and IPv6 does not remove it. A team that starts 300 agent sandboxes each second must stay below that limit. This reads as a network with more than one limit for AI systems, and a team must write down all of them before it adds more agents.
Sources
Posted