Low Latency C++ Trading System

Introduction

• Importance of building automated trading systems.
• Challenges: trading problem and technological problem.
• Increase in market participation over the last decade.

Why Low Latency is Essential

• Market makers like Optivia: numerous orders (hundreds of thousands to millions) at any given time.
• Immediate action required during market-moving events (e.g., canceling obsolete buy orders).
• Building Blocks of Low Latency Systems:
  • Data Model Access
  • System Tuning
  • Performance Measurement

Designing for Performance

• Donald Knuth: Premature optimization is the root of all evil.
• Performance should be considered at the beginning.
• Strategy vs. Tactics:
  • Strategy: Overall approach to meet performance goals.
  • Tactics: Implementation details.
• Importance of data modeling for performance.

Goals and Current Latency

• Goal: handle events with a certain latency distribution (e.g., average less than 100 nanoseconds).
• Financial market latency requirements are very stringent (e.g., ~10 nanoseconds with FPGA implementation).
• C++ still relevant for low latency due to the complexity and multiple steps in trading systems.

Data Model Access

• Profile results show that improper data modeling leads to widespread slowness.
• Example of optimizing an instrument store: using a stable vector instead of a std::unordered_map for better locality and cache efficiency.
• Importance of keeping working set size (WSS) small and localized in memory.

Concurrent Access

• Sharing Data between Applications:
  • Use of ring buffer for events.
  • State and Event approaches for data sharing.
• Example: SCC Lock (Sequence Counter Lock) for efficient read-write operations.
  • Good for systems with one producer and many consumers.
• Event Notifications using Shared Memory queues for low-latency communication.

System Tuning

• Importance of CPU Clock Frequency and Power States (C-states and P-states) for consistent performance.
• System tuning practices to minimize jitter and ensure stable performance (disable power-saving features when necessary).
• Optimization techniques include interrupt affinity, CPU core isolation, leveraging huge pages, and NUMA awareness.

Performance Measurement and Scaling

• Continual measurement and monitoring are essential for long-term system performance.
• Building performance metrics and logging into the system from the beginning.
• Example of using RdTSC for timestamping and evaluating system performance over time.

Conclusion

• Summary of key strategies and tactics discussed.
• Importance of robustness and speed in trading systems.
• References for further reading on low-latency and data modeling.

References

• Ulrich Drepper’s paper on memory and cache profiling.
• Mike Acton’s talk on data-oriented design.
• Papers on Seq Lock and performance tuning.