How to deal with the write bottleneck of mdbx under high traffic？

[fynnss]

We are doing performance stress testing for reth in a private network.

During the tests, we found that when the throughput goes beyond 3000+ TPS, reth often falls behind the head block, and switches back from live sync to stage sync, but geth can handle normal.

Our test machine is quite powerful:

• 256 GB RAM
• 24 cores

the reth-bsc(（base on reth v1.6.0）) build with the maxperf profile and the run command as above:

RUST_LOG=info ./target/release/reth-bsc node \
    --chain=xxx \
    --datadir=./data_dir \
    --authrpc.addr "0.0.0.0" \
    --authrpc.port 8551 \
    --authrpc.jwtsecret ./data_dir/jwt.hex \
    --engine.parallel-sparse-trie \
    --http \
    --http.addr "0.0.0.0" \
    --http.port 8545 \
    --metrics 0.0.0.0:6060 \
    --http.api "admin, debug, eth, net, trace, txpool, web3, rpc, reth, ots"

However, when we observed the metrics, both mdbx write latency and mdbx read latency became extremely high under load. This seems to be the main bottleneck causing the block lagging issue.

Questions

• Has anyone faced similar write throughput issues with MDBX in reth?

• Are there recommended mdbx tuning parameters or reth configurations that can help?

• Would alternative setups (e.g. tuning commit batch size, disabling sync, file system tweaks, or even alternative databases) improve performance?

Any insights, tuning experience, or best practices would be highly appreciated!