NASDAQ ITCH 5.0 Parser: 178M msg/s Single-Core on Lunar Lake

Parses all 23 NASDAQ TotalView-ITCH 5.0 message types from a real 12GB binary. Two competing implementations benchmarked against each other.

Results (Intel Core Ultra 256V, Lunar Lake, 16GB LPDDR5x)

Implementation	Single-Core	8-Thread
Hybrid dispatch	~178M msg/s	~834M msg/s
Pure branchless	~171M msg/s	~1,091M msg/s

Checksum: 0x9beeff8d334e5f70 (identical across both — proves correctness)

The Interesting Finding

Hybrid wins single core. Branchless wins multi core.

The branch predictor handles one message stream easily: Add Order ('A'), Execute ('E'), and Trade ('P') account for ~80% of real ITCH traffic, so the fast-path branches are predicted near-perfectly and cost almost nothing.

Under 8-thread load, the predictor tracks 8 independent message streams simultaneously, prediction quality degrades, and the hybrid's fast-path advantage disappears. The branchless dispatch table has no branches to mispredict and degrades gracefully.

Key Implementation Details

• mmap + MAP_POPULATE + MADV_HUGEPAGE for TLB pressure reduction
• always_inline for hot-path types (A, E, P), noinline for cold types to protect instruction cache
• __builtin_bswap + memcpy for zero UB big endian field extraction
• Thread local accumulator sink flushed per run to prove liveness without atomic contention on the hot path
• Verified with Linux perf stat TMA methodology

Build

compile any of the files with icpx and the following flags icpx -O3 -march=native -fiopenmp -flto

Data

Tested on 01302020.NASDAQ_ITCH50 (12GB). Available from NASDAQ: https://emi.nasdaq.com/ITCH/Nasdaq%20ITCH/

Perf Results

Hyrbrid Dispatch

Pure Dispatch