BENCHMARKING EMBEDDED KEY-VALUES STORES USING eBPF

Abstract Embedded key-value stores run inside latency-critical paths, yet conventional benchmarks reveal only symptoms, not causes. We present ucsb-ebpf, a drop-in extension to the Unbranded Cloud Serving Benchmark that injects eBPF probes into every run. The tracer records per-thread syscall times and optional heap/slab events while adding under 1 % runtime overhead, so the same workload that measures throughput now pinpoints where each microsecond goes. Using seven standard mixes, we profile LevelDB, RocksDB, WiredTiger and LMDB on NVMe hardware. All code, data and analysis scripts are open sourced.