Priority-fill block relay paper for 25-second blocks
zcashminer1 submitted to ZECATHON a priority-fill relay that orders block uploads to reduce orphan blocks under NU7's 25-second blocks.
The author believes that when NU7 cuts the block interval to 25 seconds on November 5 (ZIP 218), the time a block takes to cross the network becomes a larger share of the interval, so the orphan rate rises too. Today a Zebra node uploads a new block in parallel to every peer that requested it, and TCP splits bandwidth evenly, so all copies finish similarly late. The author submitted a paper and simulations on this problem to ZECATHON.
Priority fill first uploads at full speed to the peers most likely to be the next propagators, that is, those with the highest measured throughput, and uses the leftover bandwidth for the next in line. A stalled peer is ranked lower. Each node decides locally, so no new messages, consensus change or fork is needed.
On a 100-node testbed built with Zebra's zebra-network code (2MB blocks, 25-second interval), the estimated stale rate fell from 10.0% to 3.6%, and the median time to reach other miners fell from 2.67 seconds to 0.82 seconds. In a 1,000-node model using Poisson arrivals, orphan blocks fell from 16.2% to 6.0%, and the revenue gap between a 46% pool and a 1% pool narrowed from 7.3% to 2.7%.
The author also stated the limitations. Ranking must be learned, so a freshly started node gets only about a quarter of the benefit, the slowest 10% of nodes get no benefit, and attacks that manipulate the ranking are possible (10 slow attackers add a 0.8-second delay, or 0.4 seconds with a demotion rule). The model assumes full 2MB blocks, equal-hashrate miners and no compact blocks, so absolute numbers come out high and should be read for direction and rough scale. Separately, the author pointed out that zebrad's PEER_GOSSIP_DELAY (7 seconds) was set for 75-second blocks and needs to be checked for 25-second blocks. This post is for discussion, not a merge proposal.
This summary is based on the original; check the original for exact details.