Exchange | Command Routing & Partitioning

The exchange_routing module manages the concurrent topology of the exchange. It implements the "Single-Writer" pattern by partitioning instruments across dedicated worker threads.

Overview

To achieve massive throughput, the exchange does not use a single global mutex. Instead, every instrument (e.g., EURUSD) is assigned to a specific Partition. Each partition has its own lock-free command queue and a dedicated thread, ensuring that matching for one symbol never blocks matching for another.

Key Responsibilities

• Implement the Partition container lifecycle.
• Manage the WorkerThread event loop.
• Provide lock-free SPSC (Single-Producer Single-Consumer) command queues.
• Provide a unified Command interface for New, Modify, and Cancel operations.

Architecture

graph TD
    GW(Gateway) -->|submitCommand| CORE[TradingCore]
    CORE -->|Route| P1(Partition: EURUSD)
    CORE -->|Route| P2(Partition: USDJPY)
    
    subgraph "Partition Internals"
        P1 --> Q[Command Queue]
        Q --> W(Worker Thread)
        W --> MATH(exchange_matching)
    end

Class Diagram

classDiagram
    class Partition {
        +enqueue(Command*)
        +waitReady()
        -m_worker: unique_ptr~WorkerThread~
        -m_queue: spsc_queue~Command*~
    }
 
    class WorkerThread {
        -runLoop()
        -processCommand(Command*)
    }
 
    class Command {
        <<interface>>
        +getType() : CommandType
        +getOrderId() : OrderID
    }
 
    Partition *--> WorkerThread : Owns
    WorkerThread o--> Command : Processes

Component Responsibilities

Component	Description
**Partition**	A logical silo of execution. Owns all state (Book, OrderManager) for its assigned instruments.
**WorkerThread**	A high-priority system thread pinned (optionally) to a CPU core to drain the command queue.
**Command**	The common payload format used to move intentions from the networking gateway into the matching engine.

Critical Design Conventions

• Single Writer: Only one thread is EVER allowed to call exchange_matching or exchange_state for a given instrument.
• No Sleeping: In high-load scenarios, the WorkerThread can switch to a busy-wait (spin) strategy to minimize context-switch latency.