sol2go

Upgrades 2 min read

A geo-aware leader schedule: SIMD-0675 wants consecutive leaders to be neighbours

Six small white paper dots laid in an open arc across the lower half of a violet ground, the two end dots larger than the rest

Every slot on Solana has a leader, and the schedule of leaders is drawn by stake-weighted lottery. The hand-off from one leader to the next crosses whatever distance separates them, and when a validator in Frankfurt hands to one in Tokyo the light takes longer than the block. SIMD-0675, published with its companion SIMD-0674, proposes to arrange the schedule so that neighbours follow neighbours.

What it changes

Nothing about who is chosen. The proposal reorders the existing random schedule so that consecutive leaders within a run are in the same geographic region; the lottery stays stake-weighted and the set of leaders in an epoch is unchanged.

The two knobs

σ, a stake floor of about 10 per cent, below which a region is not grouped; and γ, the run length, between 3 and 7 consecutive slots given to one region before the schedule moves on.

The modelled gain

At γ = 7 the Foundation's model gives 29.1 milliseconds saved per hand-off against today's schedule. Under Alpenglow, where a block is final in about 150 ms, that is a large fraction of the budget.

The objections

The post lists them. Resilience: a regional outage during that region's run takes several consecutive slots rather than one. Asia: with about 10.4 per cent of stake, the region sits at the floor and may be grouped or not depending on the epoch. Fictitious location: a validator that lies about where it is could be scheduled into a run it then slows. The proposal is at the discussion stage.

This is the kind of proposal that only makes sense after Alpenglow. When finality was twelve seconds, thirty milliseconds of hand-off was noise; at 150 ms it is a fifth of the block. The objection that counts is the Asia one, because a floor set at ten per cent and a region at 10.4 is a rule that flips between epochs, and a schedule that is sometimes grouped and sometimes not is harder to reason about than either. Expect the number to move before the vote, and expect the proposal to pass in some form, because the people who run the largest stake are in the two regions that gain most.

A geo-aware leader schedule: SIMD-0675 wants consecutive leaders to be neighbours: questions

What does SIMD-0675 propose?

A geo-aware leader schedule: the random leader schedule is reordered so consecutive leaders in a short run are in the same region, cutting block hand-off time. Who is a leader does not change.

How much faster would Solana be?

The Foundation's model gives about 29 ms saved per hand-off at a run length of seven. It is a proposal, not an activated change.

Source: Solana Foundation checked against the source

More Solana news

A row of flat lilac paper blocks across a deep violet ground, one taller mint block among them and the blocks after it narrower and closer together, with a small cream disc high on the right

Upgrades

Alpenglow is live

Solana switched consensus from TowerBFT to Alpenglow on 21 October at slot 459,653,000, epoch 1064. Finality goes from 12.8 seconds to an expected 150 milliseconds, slots halve to 200ms, and vote transactions leave the blocks, which changes what every dashboard counts.

A pale square cut into upright strips of differing widths in mint, cream and grey-blue, standing together on a lilac ground

Upgrades

250ms slots, and smaller blocks

Solana cut its slot time to 250ms at epoch 1037 on 18 September, the fourth of five steps towards 200ms. Per-block compute fell from 60 million units to 37.5 million with it, so what this buys is latency, not capacity.

A tall stack of flat violet paper coins with the top third lifted clear of the rest

Upgrades

Rent came down. Your accounts are over-funded

Solana cut rent, so every existing account now holds more SOL than it needs. A new Token Program instruction takes the surplus out without closing anything — and this is only phase one of five.