Benchmarking chunk application on the native transfers workload

This doc shows how to:

Run a single-node localnet with native transfer workload
Use the view-state command to benchmark applying one of the chunks created during the workload

Running the workload creates ~realistic chunks and state and then we can benchmark applying one of the chunks using this.

Running the initial workload

Clone nearcore and build neard with the tx_generator feature

Note: Building neard with --profile dev-release instead of --release takes less time and the binary is almost as fast as the release one.

Note: If you plan to gather perf profiles, build with --config .cargo/config.profiling.toml (see profiling docs)

git clone https://github.com/near/nearcore
cd nearcore
cargo build -p neard --release --features tx_generator

Build synth-bm

cd benchmarks/synth-bm
cargo build --release

Go to transaction-generator

cd ../transactions-generator/

Create a link to neard

ln -s ../../target/release/neard

Set TPS to 4000

There are two files with the load schedule: tx-generator-settings.json and tx-generator-settings.json.in. Transaction generator reads data from the tx-generator-settings.json file, not the one with .in suffix. The .in file is used only to store the default values for transaction generator. If tx-generator-settings.json doesn't exist, it's created based on tx-generator-settings.json.in during the first run. This way you can modify tx-generator-settings.json locally without merge conflicts.

Create tx-generator-settings.json:

cp tx-generator-settings.json.in tx-generator-settings.json

Set tps = 4000 in tx-generator-settings.json. Example jq command (might not work in the future if the file structure changes):

jq '.tx_generator.schedule[0].tps = 4000' tx-generator-settings.json | sponge tx-generator-settings.json

Run the native transfers workload (Ignore "Error: tx generator idle: no schedule provided", this is normal during just create-accounts)

just do-it

This will create a benchmarks/transactions-generator/.near directory with node state, create accounts and run the workload. Creating accounts takes a few minutes, be patient. Sometimes running the workload fails because the database LOCK hasn't been freed in time. If that happens run just run-localnet to retry.

Re-running the workload

To re-run the workload run just run-localnet. It's not necessary to do just do-it, which would recreate the state from scratch.

To change TPS edit the tx-generator-settings.json file (without .in !) and run just enable-tx

Preparing state for chunk application benchmark

(Taken from profiling docs)

First, make sure all deltas in flat storage are applied and written:

./neard --home .near view-state --read-write apply-range --storage flat sequential

Then move flat head back 32 blocks to a height that had chunks with high load

./neard --home .near flat-storage move-flat-head back --blocks 32