Project CHRONOS

Deterministic Limit Order Book (LOB) for Sub-Microsecond Matching

A production-grade, ultra-low latency (ULL) trading system built in Java, demonstrating HFT-level performance engineering across 5 integrated services.

Architecture

┌─────────────────┐     ┌──────────────────────────────────────┐     ┌────────────────────┐
│                 │     │      Aeron Cluster (Raft)            │     │                    │
│  FIX Gateway    │────▶│  ┌─────────────┐  ┌──────────────┐  │────▶│  Response Gateway   │
│  (Ingress)      │ SBE │  │  Sequencer  │──│  Matching    │  │ SBE │  (Egress)           │
│                 │     │  │  (Spinal    │  │  Engine      │  │     │  + HDRHistogram     │
│  • TCP/NIO      │     │  │   Cord)     │  │  (Brain)     │  │     │  • P99.99 tracking  │
│  • FIX Parser   │     │  │             │  │  • SIMD/AVX  │  │     │  • SBE→FIX encode   │
│  • FIX→SBE      │     │  │  • Total    │  │  • Off-heap  │  │     │                    │
│  • Zero-alloc   │     │  │    Order    │  │    LOB       │  │     │                    │
31: └─────────────────┘     │  └─────────────┘  └──────────────┘  │     └────────────────────┘
                        └──────────────────────────────────────┘
                                         │
                        ┌────────────────────────────────────┐
                        │  Warmup Service (AOT/CDS)          │
                        │  • JIT training runs               │
                        │  • CDS archive generation          │
                        │  • GraalVM PGO support             │
                        └────────────────────────────────────┘

Key Performance Techniques

| Technique | Where | Why | |-----------|-------|-----| | Zero Allocation | FIX parser, SBE codecs, matching engine | Eliminate GC pauses — no objects created on hot path | | SIMD Vectorization | VectorizedPriceScanner | Compare 8 price levels per CPU cycle (AVX-512) | | Off-Heap Memory | OffHeapOrderBook | 64-byte cache-line-aligned order slots via Agrona UnsafeBuffer | | SOA Layout | Order book price arrays | Structure-of-Arrays enables SIMD-friendly contiguous memory access | | SBE Encoding | All inter-service messaging | Flyweight codec with zero-copy reads — no deserialization | | Aeron Cluster | Sequencer | Raft consensus for deterministic total ordering | | Fixed-Point Arithmetic | Prices | long scaled by 10⁸ — eliminates floating-point non-determinism | | Core Pinning | All services | taskset -c N eliminates context-switch jitter | | ZGC | JVM flags | Generational ZGC for sub-millisecond pause times |

Project Structure

chronos/
├── chronos-schema/          # SBE message codecs (flyweight encoders/decoders)
├── chronos-core/            # Domain model + off-heap order book
├── chronos-matching/        # SIMD matching engine (Vector API)
├── chronos-sequencer/       # Aeron Cluster service (Raft)
├── chronos-fix-gateway/     # FIX protocol ingress
├── chronos-response-gateway/# Execution report egress + latency tracking
├── chronos-warmup/          # JIT training & AOT cache generation
└── chronos-benchmarks/      # JMH + wire-to-wire + JFR zero-GC proof

Prerequisites

JDK 21+ (GraalVM CE recommended for faster JIT warmup)
Gradle 9.x (wrapper included)

Quick Start

# Build all modules
./gradlew build

# Run JMH benchmarks (SIMD vs scalar, SBE vs string-based)
./gradlew :chronos-benchmarks:jmh

# Run wire-to-wire benchmark (1M orders)
./gradlew :chronos-benchmarks:run

# Run warmup training
./gradlew :chronos-warmup:run

# Start the sequencer (Aeron cluster node)
./gradlew :chronos-sequencer:run

# Start the FIX gateway (TCP port 9876)
./gradlew :chronos-fix-gateway:run

# Start the response gateway
./gradlew :chronos-response-gateway:run

Performance Highlights

For detailed benchmark results, methodology, and the Linux optimization guide, please see BENCHMARK_GUIDE.md.

FIX Parsing: 110ns (vs 542ns QuickFIX/J)
Wire-to-Wire Latency: < 50µs (Linux Optimized)
Zero GC: Verified via JFR on critical paths.

Production Deployment

Recommended JVM Configuration

java -XX:+UseZGC \
     -XX:+ZGenerational \
     -Xms8g -Xmx8g \
     -XX:SoftMaxHeapSize=6g \
     --add-modules jdk.incubator.vector \
     --add-opens java.base/sun.misc=ALL-UNNAMED \
     -jar chronos-sequencer.jar

GC Selection:

Production: ZGC with generational mode (sub-millisecond GC pauses)
Benchmarking: Epsilon GC (no-op GC for zero-allocation verification)
Development: G1GC (default, balanced)

See docs/gc-selection-guide.md for detailed GC configuration guidance.

CPU Isolation (Linux)

# Isolate cores 2-5 from the OS scheduler
# Add to kernel boot params: isolcpus=2-5 nohz_full=2-5
taskset -c 2 java -jar chronos-sequencer.jar
taskset -c 3 java -jar chronos-fix-gateway.jar
taskset -c 4 java -jar chronos-response-gateway.jar

CDS Archive (Fast Startup)

# Step 1: Train and dump class list
java -Xshare:off -XX:DumpLoadedClassList=chronos.classlist -jar chronos-warmup.jar

# Step 2: Create CDS archive
java -Xshare:dump -XX:SharedClassListFile=chronos.classlist -XX:SharedArchiveFile=chronos.jsa

# Step 3: Run with CDS
java -Xshare:on -XX:SharedArchiveFile=chronos.jsa -jar chronos-sequencer.jar

Future Exploration (JDK 26+)

These features are currently in preview/incubator but will enhance CHRONOS when stabilized:

| Feature | JEP | Benefit | |---------|-----|---------| | Project Valhalla (Value Classes) | JEP 401 | Eliminate object headers for Order/PriceLevel — better cache density | | Project Panama (FFM API finalized) | JEP 454 | Replace Unsafe with MemorySegment + Arena for safer off-heap access | | Vector API (finalized) | JEP 489 | Remove --add-modules jdk.incubator.vector flag | | Project Leyden (AOT Cache) | JEP draft | Native AOT cache replacing CDS for instant startup |

License

Apache License 2.0

Chronos

Install / Use

README