LeiosDB benchmark CLI

The cmd/leios-db-bench command measures LeiosDB latency across fetch, vote read, serve request reads, and apply read steps.

Invocation

Run the benchmark with a supported entry point:

make bench-leios-db

Use the cached-read benchmark target to measure a read-through cache scenario. The target passes the value of LEIOS_DB_BENCH_ARGS to the benchmark:

make bench-leios-db-cached

go run ./cmd/leios-db-bench

Current flags

The latency benchmark supports the following flags:

Flag	Role
-volatile-ebs	Configures the volatile EBs used for measurement.
-request-bytes	Configures request sizing.
-job-bytes	Configures job sizing.
-upstream-peers	Sets the number of upstream peers that share fetch ranges.
-downstream-peers	Sets the number of downstream peers that share serve requests.
-samples	Sets the number of samples collected.
-gc-every	Configures the garbage collection cadence.
-burst	Enables optional burst scheduling.
-burst-size	Configures the size of scheduled bursts.
-burst-interval	Configures the interval between scheduled bursts.
-cached-reads	Enables a read-through cache for persisted manifest and closure reads for each endorser block.
-cpuprofile	Writes CPU profile diagnostics for the benchmark.
-heapprofile	Writes heap profile diagnostics for the benchmark.

Removed flags

Migration note: The following throughput benchmark flags no longer exist. An invocation that passes them fails, so existing scripts must replace them with the current flags above.

• -pre-populated-ebs
• -fetch-clients
• -ebs-per-client
• -fetch-servers
• -chain-sel-reads
• -gc-ticks
• -runs

Workload behavior

The benchmark measures one endorser block at a time across these steps:

1. Fetch reads the endorser block data.
2. Vote read reads the vote data for the block.
3. Serve request reads reads the requested data for downstream peers.
4. Apply read reads the data used by the apply step.

The benchmark partitions requests and ranges across the configured upstream and downstream peers. Pruning does not advance past endorser blocks with readers still in flight. When the benchmark ends, the optional -burst scheduler stops promptly instead of waiting for its next interval.

Output

The benchmark reports the following values for fetch, vote, serve, and apply:

• p50 latency
• p99 latency
• maximum latency
• sample count

These results describe latency for each workload step. They do not report throughput or iteration timing.

---

Docs authored by Doc Holiday