Documentation on how to get max node performance with GraalVM

[OlegJakushkin]

Description

As a PoA chain hoster, I want to understand how to increase performance using GraaLVM to get better TPS.

Acceptance Criteria

• There is a doc outlining how to run GraalVM-based besu nodes to achieve top performance
  • Possible compilation arguments/parameters
  • Possible runtime parameters and what they depend on

Logs (if a change in behaviour)

• Currently, I see such performance differences on the same node types, distributed across 3 clusters when running [Simple benchmark from Caliper] (https://github.com/hyperledger-caliper/caliper-benchmarks/tree/main/benchmarks/scenario/simple) across 2 runs with 3 validators each in physically different location:
  
• This is the container image I currently use to build and run the GraalVM besu codebase:

FROM ghcr.io/graalvm/graalvm-community:21
RUN microdnf -y install findutils 
	
COPY . /besu

WORKDIR /besu
RUN ./gradlew nativeCompile --parallel  --dependency-verification=off
RUN ./gradlew installDist  --continue --parallel   --dependency-verification=off

EXPOSE 8545 8546 8547

CMD ["/besu/build/install/besu/bin/besu", "--help"]
ENTRYPOINT ["/besu/build/install/besu/bin/besu"]

Versions (Add all that apply)

• I tried on 24.10.0 codebase

[macfarla]

@ahamlat think you looked at GraalVM recently - there were some issues with mainnet (no fallbacks for native/JNI libs) but will it work for PoA? Any insights on this?

[ahamlat]

@OlegJakushkin What documentation are your referring to and what do you mean by GraalVM-based besu nodes ?
As Besu is written in JAVA, you can change the JVM implementation to choose the GraalVM one.
For the EVM module, we can execute som tests with a native GraalVM image. I started a work to create a full native image for Besu but had faced some challenges, as all the libraries used in Besu are not supported yet by GraalVM, so debugging one by one.

[OlegJakushkin]

Dear @ahamlat:

1. Is there an option to run the main besu client with EVM compiled as a native part (what am I missing in building a pipeline in the docker file above, or maybe there is some node running option to pass it for native EVM to be used by the client)?
2. In here, it was posted that on graalvm-based client shall have +50 tx/s in comparison with OpenJDK from roughly 350 to 400 smart contract calls per second so I wonder if there are any build/runtime args to get that speed up as so far with same args I get a slowdown in my use case.