citadel_skill

Overview

This skill captures Citadel Securities–style engineering patterns for building ultra-low-latency, deterministic trading systems. It emphasizes kernel bypass networking, lock-free data structures, CPU isolation, and zero-allocation hot paths to minimize jitter and tail latency. Use it as a practical checklist and implementation guide when designing latency-critical market making or execution engines.

How this skill works

The skill inspects architectural decisions and provides concrete patterns: kernel-bypass packet I/O (DPDK/OpenOnload), pre-allocated object pools, lock-free order books, CPU pinning and isolation, and nanosecond-precision latency measurement. It translates those principles into concrete code patterns, system configuration steps, and operational rules oriented around minimizing worst-case latency and eliminating nondeterminism.

When to use it

• Building market-making engines or high-frequency trading infrastructure
• Designing any latency-critical data path where microseconds matter
• Replacing blocking I/O or mutex-based hot paths with deterministic alternatives
• Hardening production systems against tail-latency and jitter issues
• Profiling and validating latency budgets end-to-end

Best practices

• Measure latency at every component boundary and budget nanoseconds per component
• Use kernel bypass (DPDK/OpenOnload) and eliminate syscalls on the hot path
• Pin critical threads to dedicated cores, set real-time priority, and isolate from the scheduler
• Pre-allocate memory and use object pools to avoid runtime allocation or page faults
• Favor lock-free data structures, seqlocks, and busy-polling over blocking in hot paths
• Verify compiled output on critical loops and prefer fixed-size, cache-aligned data layouts

Example use cases

• High-frequency market maker that processes market data and emits orders with sub-microsecond consistency
• Market data receiver using DPDK-style burst polling and prefetching for minimal packet handling latency
• Lock-free top-of-book snapshot readers for many concurrent consumers with a single writer
• Pre-faulted object pool for order and message objects to achieve zero-allocation hot paths
• System boot and kernel tuning scripts to isolate CPUs, disable C-states and THP for deterministic performance