deterministic subsetting algorithm to limit the number of instances

[jizhuozhi]

Is your feature request related to a problem? Please describe.
Many RPCs use connection pools and long connections to reduce TCP handshake overhead (including RPC over HTTP). In a very large-scale cluster, there may be tens of thousands of instances (such as a search engine). If the client establishes connections with all instances, all available TCP ports will be quickly exhausted.

Describe the solution you'd like
In order to solve this problem, in very large-scale clusters, each client should only select a limited subset to establish connections. Google SRE provides a deterministic subsetting algorithm solution, see https://sre.google/sre-book/load-balancing-datacenter