Anza and a16z Researchers Publish Gatling: A Protocol Achieving 10ms Slots and 214ms Transaction Latency on Solana

Researchers from Anza and a16z Crypto Research published a new consensus protocol paper on June 16, 2026, showing that Solana-compatible block proposal times can shrink from 400 milliseconds to around 10 milliseconds by running multiple consensus instances simultaneously rather than sequentially.

The paper, titled "Gatling: Rapid-Fire Consensus from Parallel Composition", was authored by Giulia Scaffino (TU Wien, Common Prefix), Max Resnick (Anza), and Joachim Neu (a16z Crypto Research). Scaffino and Resnick announced the work publicly on June 17, with Solana co-founder Anatoly Yakovenko retweeting Resnick's thread.

The Leader Handoff Bottleneck

Current Solana validators operate under a rotating-leader schedule where each designated leader gets roughly 400 milliseconds to produce a block before the next leader takes over. The handoff creates a structural delay: the incoming leader must receive the previous leader's block before it can begin proposing its own. Under realistic wide-area network conditions, that wait alone can consume most of the slot budget.

Gatling removes this dependency by running multiple independent consensus instances in parallel, each with its own leader schedule, staggered in time. Rather than one instance proposing every 400ms, nine parallel instances each propose at 500ms intervals, offset so that a new proposal arrives roughly every 56ms. A deterministic interleaving rule then merges the outputs from all instances into a single ordered log, preserving consistency without requiring any instance to wait for another before proposing.

The paper frames this as "black-box composition": Gatling wraps any existing atomic broadcast protocol (the prototype uses threshold-Simplex) without modifying its internals. The composition alone is what drives the speedup.

Gatling Prototype Results: 226ms Latency on a Global 10-Node Cluster

The Anza research team tested the prototype on a 10-node globally distributed cluster (the same "Anza invalidator cluster 1" used for internal benchmarks) with nodes in London (1), Tokyo (3), Singapore (3), Dallas (2), and Miami (1). Network delays between the most distant nodes reach approximately 110–112 milliseconds between Miami and Singapore.

With nine parallel instances (K=9) and 500ms component inter-proposal times, the prototype achieved end-to-end transaction latency of 226.6ms under clean network conditions, compared to current Solana slot times above 400ms. The slot inter-proposal time measured approximately 56ms, roughly a factor-of-nine reduction from the 500ms single-instance baseline. The skip rate held at 0.3% under normal conditions; at 1% packet loss, latency rose to 244ms.

The 10ms slot figure cited in the announcements reflects the theoretical minimum at the paper's optimal configuration, where the inter-proposal time ε equals 2Δ/K with Δ being network delay and K the number of parallel instances. At K=9 with tightly controlled network conditions, this approaches 10ms. The 226–244ms end-to-end latency measures how long a transaction takes from submission to finality across the global cluster.

How Gatling Handles Crashed Leaders and Head-of-Line Blocking

Parallel instances introduce a new complication: if one leader crashes, its instance stalls. Because the interleaving rule requires all prior slots across all instances to be decided before later blocks can be appended, a crashed leader creates a head-of-line block that delays confirmation of blocks from healthy instances.

Gatling handles this with an empty-blocks variant: stalled instances confirm empty blocks through the normal consensus mechanism, releasing the log jam. The paper analyzes this case formally and derives an optimal K for a given success probability per proposal attempt. At a 99% per-attempt success rate, the theoretical minimum latency approaches approximately 3.09 times the base confirmation delay, a result the authors show improves upon highly optimized single-instance protocols.

Relationship to Constellation and Alpenglow

Resnick was explicit in his announcement that Gatling is research and does not represent a change in Anza's current development direction. "This is just research and doesn't represent a change in our direction with respect to Constellation," he wrote. "This type of architecture is orthogonal to multiple proposer designs."

Constellation, Anza's SIMD-0322 proposal for concurrent block producers, addresses who proposes blocks. Gatling addresses how fast a single rotating-leader schedule can propose them. The two are combinable: a Gatling-style parallel composition can run on top of a Constellation-style multi-proposer foundation.

The paper also cites Alpenglow, Anza's in-progress consensus upgrade targeting 150ms finality, as an example of a slowly-rotating-leader protocol that Gatling's approach would complement.

From Protocol Theory to Practice

The Gatling prototype is built in Rust on top of Commonware's "alto" blockchain codebase and is published on GitHub under a research license. The repository's readme is explicit: "This code is for research purposes only and it should not be used in production."

The paper closes with a finding that "parallel composition alone is sufficient to achieve latencies improving upon those of highly optimized protocols" even when starting from an unoptimized baseline, which the authors argue makes Gatling a useful building block for future protocol design regardless of which underlying consensus algorithm a chain uses.

Anza has not announced a timeline or development plan for integrating Gatling into the Solana validator client.