Orderbook Implementations

Testing different orderbook implementations with live market data.

1. Map - std::map
2. Vector - sorted vector with binary search
3. Vector (reversed) - reversed for better access patterns
4. Vector (branchless) - branchless binary search
5. Vector (likely) - compiler hints
6. Vector (linear) - linear search

Results

Benchmarked on live BTC-USD data from Coinbase (100K warmup, 500K test):

Implementation	P50	P90	P99	P99.9	Mean
Map	46 ns	99 ns	181 ns	306 ns	51 ns
Vector	10 ns	17 ns	25 ns	70 ns	11 ns
Vector (reversed)	11 ns	18 ns	27 ns	57 ns	12 ns
Vector (branchless)	10 ns	17 ns	33 ns	51 ns	12 ns
Vector (likely)	10 ns	15 ns	33 ns	42 ns	11 ns
Vector (linear)	80 ns	93 ns	100 ns	167 ns	84 ns

Branch Prediction

perf stat results on 50K modify operations:

Standard Binary Search:

• Branches: 36.5B
• Branch misses: 44.1M (0.12%)
• Cycles: 70.1B
• Instructions: 165.5B

Branchless Binary Search:

• Branches: 19.3B (47% fewer)
• Branch misses: 28.7M (0.15%)
• Cycles: 37.3B (47% fewer)
• Instructions: 87.6B (47% fewer)

Branchless doesn't predict branches better, it just removes them entirely.

Structure

src/             orderbook implementations + coinbase client
include/         headers
bench/           benchmarks (coinbase_live_bench, branch_test)
tests/           unit tests