I like the way Ripple provides effectively instantaneous transaction verification.
When building consensus how much traffic and their equivalent of orphan blocks are accumulated?
Is communication throttled at all?
Is the candidate set processed by the network or the founders?
When building consensus how much traffic and their equivalent of orphan blocks are accumulated?
Once a consensus round starts, the amount of information that needs to be distributed is fixed, since new transactions cannot be introduced. If there's no reason a transaction shouldn't be included, then almost all honest nodes will ask that it be included. If there's any reason a transaction shouldn't be included, then the avalanche algorithm quickly excludes it.
Is communication throttled at all?
Each validator rate-limits how often it changes its position. There is no reason to send a position that is already obsolete due to the decisions made by other validators.
Is the candidate set processed by the network or the founders?
Once each validator determines that a particular candidate set has a super-majority, it applies the transactions in that candidate set and signs the resulting ledger. If a super-majority of validators sign that ledger, then we have a new "last fully-validated ledger".
Any transactions that could have gotten into that candidate set but didn't are then recovered using a deterministic process.
how much time does the average round consume?
The network bases each round's timing on the previous rounds, trying to make the round a little bit faster if it can, but making it slower if it must.
There are a few minimum timings to ensure network stability. A ledger must be open for at least two seconds in order to ensure nodes have a chance to finish processing the previous ledger. A consensus cannot be declared for at least two seconds after the ledger closes to ensure nodes have a chance to take an initial position before other nodes conclude there is a consensus. (These are absolute minimums. If needed, they will rise.)
This makes a typical consensus round take about 5 seconds, plus 1 second for each additional avalanche needed to reach consensus. The number of rounds tends to grow logarithmically with the number of nodes and linearly with the network diameter (time it takes for a flooded message to reach the majority of nodes). We don't expect the timing to get over 10 seconds, even with a large network, so long as the network is confined to Earth.
Underperforming nodes will "bow out" of the consensus process to prevent the network from becoming as slow as the slowest node. Overperforming nodes will slow down to prevent the network from becoming as fast as the fastest nodes (as this would reduce decentralization). The algorithm is designed to self-regulate to keep the fastest subset of sufficient size operating.
