How to Build a High-Performance Blockchain

Development
,
1

By Zhuolun Xiang, Siyuan Han, Zekun Li, and Alexander Spiegelman

:rocket: Developer-Focused Summary

In this paper, Xiang, Han, Li, and Spiegelman explore how to engineer high-performance blockchains by applying pipelining principles from hardware, databases, and AI systems. Their core message is clear: to achieve both high throughput and low latency, blockchain design must move beyond sequential processing and embrace concurrent, parallelized architectures.

:building_construction: The Evolution of Blockchain Pipelines

:cross_mark: Legacy Design: Sequential Execution

  • Blocks are processed one at a time: proposal → execution → commit → repeat.
  • Validators often sit idle, leading to poor resource utilization and high end-to-end latency.

:white_check_mark: Aptos: The Ford Factory of Blockchains

Aptos introduces a pipelined blockchain architecture, inspired by assembly lines:

  • Asynchronous Execution & Commit (DIP-213): Consensus first, execution follows later.
  • Pipelined Consensus (Jolteon): Start new consensus rounds before the current one finishes.
  • Quorum Store: Offloads data dissemination from consensus to all validators, avoiding leader bottlenecks.

:hammer_and_wrench: Dev Insight: Aptos turns validator resources into a parallel pipeline, maximizing throughput by running different stages of different blocks simultaneously.

:high_voltage: Zaptos: Latency-Optimized Blockchain Design

Zaptos goes further—targeting sub-second latency without sacrificing throughput by embedding pipeline stages inside consensus:

  • Optimistic Execution: Execute blocks as soon as they’re proposed.
  • Optimistic Commit: Persist state pre-certification, revert if needed.
  • Fast Certification: Begin certifying blocks before consensus completes.

:brain: Impact for Engineers:

  • Pipeline latency becomes nearly equal to consensus latency.
  • At 20K TPS, latency drops from 1.32s (Aptos) to 0.78s (Zaptos).
  • Full node and validator workloads are parallelized and overlapped.

:bar_chart: Benchmark Results

  • Test Setup: 100 validators, 30 fullnodes, 10 cloud regions (Google Cloud)
  • Latency: Zaptos maintains sub-second confirmation even under heavy load
  • Throughput: Stable at 20,000 transactions per second (TPS)

:toolbox: Key Takeaways for Blockchain Developers

  • Design for concurrency: Run consensus, execution, commit, and sync in parallel.
  • Minimize latency with speculation: Start execution and commit optimistically.
  • Avoid centralized bottlenecks: Use leaderless or horizontally-scaled data dissemination.
  • Shadow latency inside consensus: Compress critical paths for real-time user experience.
4 Likes