Twelve exchanges, one normalized tape, sub-10-microsecond wire-to-book. A co-located feed handler that turns raw multicast chaos into a single clean stream, at 20 million messages a second.
Every microsecond is accounted for before a line of logic runs. DPDK removes the kernel from the data path, hugepages eliminate TLB misses, and the FPGA decodes exchange protocols in hardware so the CPU never spends cycles parsing bytes.
The C++ book engine is the only software in the hot loop: lock-free, zero-allocation, branch-free updates on cache-aligned structures, pinned to isolated cores with hyperthreading and power management disabled. Jitter is the enemy, so the whole node is tuned to make the fast path the only path.
Each exchange ships its own binary protocol, sequencing rules, and A/B redundant feeds. The normalizer arbitrates the two lines, detects sequence gaps, requests retransmits, and emits a single canonical book event, so downstream code never sees an exchange-specific quirk.
That uniformity is what makes the replay store valuable: the same code path serves live trading and 5-year historical backtests, byte-for-byte, with no live-versus-replay divergence to debug.
| Component | Technology | Why |
|---|---|---|
| Packet ingest | DPDK + hugepages | Kernel bypass kills syscall and copy overhead |
| Line decode | FPGA | Deterministic parse + hardware wire timestamps |
| Book engine | C++ lock-free | Branch-free, cache-aligned, zero-allocation hot path |
| Tick store | kdb+ | Column store built for time-series at scale |
| Clock sync | PTP / PPS | Nanosecond cross-feed ordering |
| OS / runtime | Tuned Linux |
DPDK poll-mode drivers read multicast directly from the NIC with zero kernel involvement.
FPGA stamps every packet at the wire with PTP-disciplined clocks for true cross-feed ordering.
Order books rebuild via branch-free, zero-allocation updates that stay resident in cache.
Dual redundant lines are merged with gap detection and automatic retransmit recovery.
A tickerplant fans the unified stream to subscribers and persists it for in-memory analytics.
Five years of journaled ticks replay through the identical hot path used in production.
“If the kernel touches your packet, you have already lost the trade.”
The result is a single co-located node that ingests twelve raw exchange feeds, normalizes them into one canonical tape, and delivers sub-10-microsecond wire-to-book latency at 20 million messages per second, with every tick durably journaled for tick-for-tick 5-year replay. By pushing protocol decode into an FPGA, bypassing the kernel with DPDK, and keeping the only software in the loop lock-free and cache-resident, the platform gives trading and research teams one clean, deterministic stream that behaves identically live or in backtest.
| Isolated cores, no NUMA crossings, no jitter |