When Nanoseconds Matter: Ultrafast Trading Systems in C++

on YouTube ↗01h 29m

" summary

Claims are concrete: an order book with ~1000 levels per side is fastest as a contiguous array reversed so top‑of‑book updates avoid vector shifts, linear search beats binary on hot short ranges and branchless lower_bound raises IPC from ~1.4 to ~1.6. Tools are real: perf, hardware counters, clang x‑ray, AMD Z4 Zen4 microbenchmarks, Solarflare/onload and efvi/DPDK for networking. Numbers that surprise: kernel UDP ≈3 μs vs ≈700 ns layer‑2 UDP, 73‑byte messages for queue tests. Built pieces: a simple 150‑line shared‑memory bounded queue with reservation trick (reserve 100 KB) that reduces atomic contention and measures well.

AI-written summary · about

" chapters (37)

1. 0:00Ultrafast trading systems intro
2. 1:16Roman Empire lessons on planning
3. 4:50Market making as losers game
4. 7:57Software still matters in low latency
5. 10:45Order book fundamentals and data structures
6. 15:39Order book with std::map limitations
7. 17:23Latency distribution of data structures
8. 20:11Avoid node-based containers
9. 21:57Replace std::map with std::vector
10. 25:50Understand domain for half-solved problems
11. 26:55Leverage problem-specific properties
12. 27:16Profiling with perf and Top-Down Analysis
13. 30:07Branch mispredictions in binary search
14. 31:51Branchless binary search optimizations
15. 33:37Simplicity as ultimate optimization
16. 35:44Mechanical sympathy with hardware
17. 36:47Instruction cache and inlining trade-offs
18. 39:16Lambda optimizations pitfalls
19. 40:19Networking and concurrency in low latency
20. 41:23Bypass kernel for ultra-low latency
21. 42:25Networking APIs performance comparison
22. 43:07Mindful use of system components
23. 44:29Shared memory fundamentals
24. 45:56Bounded concurrent queues design
25. 49:11Right concurrency tool choice
26. 51:21Fast concurrent queue with atomics
27. 53:06Data races and atomic operations
28. 54:35Concurrent queue benchmarks
29. 56:01Atomic counter batching optimization
30. 1:01:16Intrusive profiling for bottlenecks
31. 1:04:43Clang X-Ray for production profiling
32. 1:06:29Stay fast with constant vigilance
33. 1:07:56System-wide performance awareness
34. 1:09:20Cache hierarchy impact on workers
35. 1:13:13No silver bullets in low latency
36. 1:14:39References and further learning
37. 1:15:00Q&A on trading system optimizations

" related talks

• 47mBenchmarking C++, From video games to algorithmic tradingAlexander Radchenko · Meeting C++ 2018
• 01h 00m“Want fast C++? Know your hardware!"Timur Doumler · CppCon 2016
• 59mCounting Nanoseconds Microbenchmarking C++ CodeDavid Gross · Meeting C++ 2019
• 40mLudicrous Speed: Designing For Performance on the JVMCliff Moon · Strange Loop 2014
• 33mReaching Ludicrous SpeedMatteo Collina · Codemotion Amsterdam 2016
• 59mPerformance matters - new tricks for old dogsMatt Bentley · Meeting C++ 2023
• 01h 04mDesign for PerformanceFedor Pikus · CppCon 2018
• 01h 38mKeynote: Benchmarking - It's About TimeMatt Godbolt · C++Now 2026