Research

RISE PEVM: Parallel EVM

By Sam Battenally

Sam Battenally

03 Jul 2024 • 3 min read

RISE Labs

We are RISE Labs, an R&D company setting out to build the most performant EVM rollup stack.

After several months of building in stealth mode, we are excited to start sharing our work with the community. This first article focuses on our open-source Parallel EVM engine, which massively speeds up block syncing and building time.

Why Parallel EVM?

In April, the cumulative activity of all Ethereum rollups ranged between 100 and 200 TPS, this is across 50+ rollups! Compare that to Solana, which consistently produces 1000–2000 TPS, albeit with caveats, it’s ~10x all rollups combined. So how do we speed up the EVM?

For a long time DA was the bottleneck, that’s no longer the case thanks to EIP4844 and external DA options. The new frontier is execution. We need to process EVM blocks faster. That simple.

What is PEVM

RISE pevm is EVM execution on steroids. It’s a parallel EVM implementation, compatible with reth. EVM blocks are inherently sequential, so typically transactions are executed sequentially. However, pevm allows us to split the task of executing transactions across many CPU cores, while maintaining a deterministic outcome.

PEVM is pre-alpha, however we can share the initial benchmarks. See here for specific details about the benchmarks.

The first table details results for a benchmark for large blocks with few dependencies. This illustrates the potential of parallel EVM when fully harnessed. We see that Uniswap swaps perform the best, with a 10.55x improvement; intuitively, this is because swaps require more computation than transfers. But how does it perform with typical real-world blocks?

A variety of Ethereum blocks were chosen randomly and tested. The key result is pevm is 1.73x faster on average than the typical sequential execution. The fastest block was 3.62x faster, but some block were slower. This is typical for the smaller blocks with more dependencies that we see in early Ethereum history.

These are the two extremes, ideal big blocks vs typical Ethereum small blocks. For L2’s with large blocks, we anticipate pevm will consistently surpass 5x. More testing required!

Note: These tests are conducted in a closed environment, we’re testing the execution stage purely with an in-memory state. Mileage will vary when running a full node, state access and additional steps in the block processing pipeline will bring these numbers down significantly.

Where We Are

Our working version passes the Ethereum General State Tests and comfortably executes Ethereum Mainnet blocks. Our Alpha, set for mid-June, is only blocked by a rare race condition and a more stable API.

Next steps

pevm is already the Fastest Execution option available, we’re looking forward to running our first reth sync with pevm to share the significant speed up.

We expect to improve another 3~5x with further optimizations. Most notably:

Optimize concurrent data structures to maximize CPU cache and stack memory.
More granular memory locations (like breaking AccountInfo down into balance, nonce, and code) to avoid false positive dependencies.
Add pre-provided metadata from a statically analysed mempool or upstream nodes.
Write custom memory allocators for the whole execution phase and the multi-version data structure. Early experiments with jemalloc, mimalloc, and snmalloc show potential up to 50% improvements. This is understandable when we optimize to the microseconds.
Support multiple EVM executors (REVM, JIT & AOT compilers, etc.).

And more!

Acknowledgements

reth & revm provided a solid baseline and flexibility to enable pevm. Many thanks to Paradigm for their solid work on both.

Our Mission

At RISE, we are committed to solving all performance bottlenecks to scale the EVM chains we love. We are planning more articles on Parallel EVM and state access, so stay tuned.

Join us to Realise Internet Scale Ethereum!