Performance & JMH Benchmarks (v1.0.0)

TangoDB uses the Java Microbenchmark Harness (JMH) to empirically measure, analyze, and validate engine throughput, native allocation overheads, and disk serialization speeds.

---

Benchmark Results Summary (v1.0.0)

The table below reports raw JMH microbenchmark throughput scores (ops/s) captured from benchmark runs stored in benchmark_screenshot/v1.0.0/jmh_benchmark_results.txt.

Benchmark Class & Method	Parameter (records)	Mode	JMH Score (ops/s)	Error Margin	Benchmark Operation Scope
MemtablePutBenchmark.put	N/A	Throughput	40,769,881.228	± 24,019,114.386	1 Single Key-Value Record Insert (~24.5 ns/op)
TangoPutBenchmark.put	N/A	Throughput	1,545.959	± 500.907	1 Full End-to-End Put (WAL Sync + MemTable)
SstManagerReadBenchmark.read	N/A	Throughput	512.502	± 3,871.974	1 Indexed SSTable Point Read (Bloom Filter + Sparse Index)
SstManagerWriteBenchmark.write	1,000	Throughput	447.738	± 1,458.170	1 Complete SSTable Flush (1,000 records)
SstManagerWriteBenchmark.write	10,000	Throughput	290.835	± 66.841	1 Complete SSTable Flush (10,000 records)
SstManagerWriteBenchmark.write	100,000	Throughput	59.042	± 1.272	1 Complete SSTable Flush (100,000 records)

---

1. MemTable In-Memory Ingest Benchmark

MemtablePutBenchmark.put measures single-threaded key-value put() operations into off-heap MemTable memory segments managed via Java's Foreign Function & Memory (FFM) API (MemorySegment and Arena).

Execution Details & Environment

• JMH Harness Version: 1.37
• JVM Runtime: OpenJDK 21 (64-Bit Server VM, Oracle Corp JDK 21.0.8+12-LTS-250)
• OS / Environment: Windows AMD64
• Flags: --enable-preview --enable-native-access=ALL-UNNAMED

Raw JMH Output

Benchmark                  Mode  Cnt         Score          Error  Units
MemtablePutBenchmark.put  thrpt    3  40769881.228 ± 24019114.386  ops/s

• JMH Score: 40,769,881 ops/sec (~40.77 Million single key-value record puts per second)
• Latency: ~24.5 nanoseconds per operation

---

2. SSTable Indexed Read Benchmark (Bloom Filter & Sparse Index)

SstManagerReadBenchmark.read measures point read throughput across .sgdb SSTable files on disk.

Architectural Drivers

1. O(1) Fast Negative Rejection: Embedded a 1% false-positive rate Bloom Filter (Bfilter). Queries for non-existent keys are rejected in nanoseconds without reading disk blocks.
2. O(log(N/16)) Binary Search: Sampled Sparse Index entries every 16 records allow readers to binary search key intervals and jump directly to the target 16-record chunk on disk.
3. Positional Reads: Refactored SstReader to use positional FileChannel.read, completely eliminating native OS mmap memory leaks and Windows file locking retention.

Benchmark                      Mode  Cnt    Score      Error  Units
SstManagerReadBenchmark.read  thrpt    3  512.502 ± 3871.974  ops/s

---

The Optimization Journey

  Phase 1: Heap Allocations        Phase 2: FFM Native Segments      Phase 3: Bloom Filter & Sparse Index
┌──────────────────────────┐    ┌──────────────────────────┐    ┌──────────────────────────┐
│ On-heap allocations      │───►│ Move record payload      │───►│ Pre-allocate 64 MiB      │
│ induced GC pauses under  │    │ off-heap via FFM         │    │ Arenas, ThreadLocal key  │
│ high write throughput.   │    │ MemorySegments.          │    │ lookups & Sparse Index.  │
└──────────────────────────┘    └──────────────────────────┘    └──────────────────────────┘
                                                                             │
                                                                             ▼
                                                                  40.7M+ ops/sec MemTable Put
                                                                  512.5 ops/sec Indexed SSTable Read