SIMD-0675 draft: a geo-aware leader schedule would cut mean handover latency from 36 to 17 ms
Serguey Asael Shinder2 min readA draft Solana Improvement Document, SIMD-0675 "Geo-Aware Leader Schedule", proposes changing how each epoch's leader schedule is built. Authored by Quentin Kniep and Roger Wattenhofer and opened on 29 September, the pull request was updated again on 15 October; its status in the document is Draft.
The idea. Today the schedule is a stake-weighted random order (SIMD-0180). SIMD-0675 adds a post-processing step that reorders that schedule so that geographically close leaders follow each other in short runs. It does not change how many leader windows anyone gets: every validator keeps exactly as many as in the base schedule. Locations come from validators' self-reported positions under a companion proposal, SIMD-0674.
Why it matters under Alpenglow. With Alpenglow's fast leader handover (SIMD-0630), each leader sends its block directly to the next one, so the delay between two leaders is dominated by the one-way network latency between them. The authors name two problems with a random order: intercontinental hops between consecutive leaders, which they call the main remaining reason an epoch runs longer than its nominal duration, and a centralisation incentive — a validator in a stake-dense location is, on average, close to its predecessor and receives blocks earlier, which rewards co-location and penalises remote operators.
The constants. The draft sets RUN_LENGTH = 3 and STAKE_FLOOR = 10%. Each validator gets a radius: the smallest distance within which at least 10% of total stake sits. Bins of nearby leaders may only form within that radius, which sets a minimum diversity per bin and limits how easily a low-stake attacker can fill an entire run. Longer runs cut latency further but mean more time in a single jurisdiction, more correlated failures and longer runs for a potential adversary.
The simulation. On the mainnet stake distribution of epoch 1038 — 661 validators, locations corrected with Globalping measurements, 108,000 leader windows per epoch, five random seeds — the mean handover latency between honest validators falls from 36.2 ms today to 17.0 ms with runs of three; the median over all handovers falls from 23.4 ms to 4.5 ms. The authors model a near worst-case attacker with 5% of stake placed alone in Sydney, 73 ms from the nearest stake in Los Angeles. In today's schedule 94.7% of its runs have length one; with RUN_LENGTH = 3, 53.5% are length one, 24.0% length two and 22.3% length three.
A knock-on change. Because handovers get shorter, the draft says HANDOVER_COMPENSATION from SIMD-0630 must rise from 25 ms to 50 ms on activation: in the simulation the average time from receiving the previous leader's block to issuing the ParentReady event grows from 23.2 ms to 46.2 ms. The authors add that this value should be revisited as the geography of stake shifts. No feature gate or activation date is set; the feature field in the document is still a placeholder.
Source: SIMD-0675 pull request — https://github.com/solana-foundation/solana-improvement-documents/pull/675