This directory contains the upstream C benchmark provided by the LMDB maintainer in response to ITS#10406 - our report of write performance regression.

This benchmark serves as an authoritative baseline for validating LMDB performance across versions:

• Uses native C code (eliminates JNI overhead)
• Provided directly by LMDB maintainer
• Matches our Java benchmark scenario (sequential append with integer keys)
• Enables comparison between Java and C performance characteristics

• mtest-append-original.c - Upstream benchmark code from ITS#10406 (unmodified)
• mtest-append.c - Modified benchmark for testing different scenarios
• run-upstream-bench.sh - Test harness to run across multiple LMDB versions
• target/ - Build artifacts and results (gitignored)
  • target/mtest-append - Compiled benchmark binary
  • target/results/ - Benchmark results for each LMDB version

• Operation: Sequential writes using MDB_APPEND flag
• Key Type: Integer keys (MDB_INTEGERKEY)
• Entry Count: 1,000,000
• Value Size: 100 bytes (zeroed)
• Commit Frequency: Every 1,000 entries (1,000 transactions total)
• Flags: MDB_NOSYNC (for consistency with Java benchmarks)
• Iterations: 3 runs per version

./run-upstream-bench.sh

Results are saved to target/results/upstream-LMDB_<version>.txt

The script automatically:

1. Checks out each LMDB version from the OpenLDAP repository
2. Compiles the LMDB library
3. Compiles and runs the benchmark
4. Generates a summary comparison table

Each benchmark iteration completes in ~0.075 seconds on modern hardware. Total runtime for all 14 versions (3 iterations each): ~1 minute.

This C benchmark can be compared with LmdbJava benchmarks at https://lmdb-benchmark.lmdbjava.org/

Key differences:

• C benchmark uses raw LMDB API (no JNI overhead)
• C benchmark commits every 1,000 entries vs Java's single-transaction approach
• Results help isolate whether regressions are in LMDB itself or JNI layer

Licensed under the OpenLDAP Public License (see mtest-append-original.c for copyright details).