https://github.com/Galxe/grevm
Parallel EVM inspired by BlockSTM
https://github.com/Galxe/grevm
evm parallel-evm reth
Last synced: 9 months ago
JSON representation
Parallel EVM inspired by BlockSTM
- Host: GitHub
- URL: https://github.com/Galxe/grevm
- Owner: Galxe
- License: apache-2.0
- Created: 2024-10-14T08:41:37.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2025-08-29T05:29:42.000Z (10 months ago)
- Last Synced: 2025-09-13T10:32:54.905Z (9 months ago)
- Topics: evm, parallel-evm, reth
- Language: Rust
- Homepage: https://docs.gravity.xyz/research/grevm
- Size: 2.14 MB
- Stars: 32
- Watchers: 5
- Forks: 12
- Open Issues: 8
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
- awesome-reth - grevm
README
# Grevm
> Grevm (both 1.0 and 2.1) is reth-ready, please see [use-with-reth.md](docs/use-with-reth.md) for more details.
Grevm is a Block-STM inspired optimistic parallel EVM execution engine that leverages DAG-based task scheduling, dynamic
dependency management, and parallel state storage to significantly boost throughput of
[revm](https://github.com/bluealloy/revm), while reducing CPU overhead in high-conflict scenarios.
## **TL;DR – Highlights of Grevm 2.1**
- **Grevm 2.1 achieves near-optimal performance in low-contention scenarios**, matching Block-STM with **11.25
gigagas/s** for Uniswap workloads and outperforming it with **95% less CPU usage** in inherently non-parallelizable
cases by **20–30%**, achieving performance close to sequential execution.
- **Breaks Grevm 1.0’s limitations in handling highly dependent transactions**, delivering a **5.5× throughput
increase** to **2.96 gigagas/s** in **30%-hot-ratio hybrid workloads** by minimizing re-executions through **DAG-based
scheduling** and **Task Groups**.
- **Introduces Parallel State Store**, leveraging **asynchronous execution result bundling** to **overlap and amortize
30-60ms of post-execution overhead within parallel execution**, effectively hiding these costs within execution time.
It also seamlessly handles **miner rewards and the self-destruct opcode** without the performance penalties of
sequential fallbacks.
- **In-depth analysis of optimistic parallel execution** reveals the **underestimated efficiency of Block-STM** and the
strength of **optimistic parallelism**, providing new insights into parallel execution.
- **Lock-Free DAG** (introduced in 2.1) replaces global locking with fine-grained, node-level synchronization. This
change reduces DAG scheduling overhead by **60%** and improves overall performance by more than **30%**. In workloads
with fast-executing transactions—such as raw and ERC20 transfers—it delivers nearly **2×** higher throughput.
## Architecture Overview
Grevm 2.1 is composed of three main modules:
- **Dependency Manager (DAG Manager):**
Constructs a directed acyclic graph (DAG) of transaction dependencies based on speculative read/write hints.
- **Execution Scheduler:**
Selects transactions with no dependencies (out-degree of 0) for parallel execution, groups adjacent dependent
transactions into **task groups**, and dynamically updates dependencies to minimize re-execution.
- **Parallel State Storage:**
Provides an asynchronous commit mechanism with multi-version memory to reduce latency and manage miner rewards and
self-destruct opcodes efficiently.
## Running the Benchmark
To reproduce the benchmark:
```bash
JEMALLOC_SYS_WITH_MALLOC_CONF="thp:always,metadata_thp:always" \
NUM_EOA= HOT_RATIO= DB_LATENCY_US= \
cargo bench --bench gigagas
```
Replace ``, ``, and `` with your desired parameters.
## Further Details
For a comprehensive explanation of the design, algorithmic choices, and in-depth benchmark analysis, please refer to the
full technical report.
- [Grevm 2.1 Tech Report](docs/v2/grevm2.1.md)
- [Grevm 1.0 Tech Report](docs/v1/README.md)