How Stock Exchanges Work: Low-Latency Trading System Design

The Need for Speed

Stock exchanges are among the most demanding systems in existence. The NYSE processes over 3 billion messages daily with latency requirements in microseconds - not milliseconds, microseconds. Understanding this architecture reveals the extreme end of system design.

Core Components

1. Order Gateway

• Entry point for orders
• Validation and normalization
• Risk checks
• Rate limiting

2. Matching Engine

• Heart of the exchange
• Matches buy and sell orders
• Executes trades
• Maintains order books

3. Market Data Publisher

• Distributes price updates
• Multicast for efficiency
• Sub-millisecond delivery

4. Risk Management

• Pre-trade risk checks
• Position limits
• Circuit breakers

The Matching Engine

Order Book Structure:

Each security has an order book:

• Bid side: buy orders (highest price first)
• Ask side: sell orders (lowest price first)
• Price-time priority within price levels

Matching Algorithms:

1. Price-Time Priority (FIFO)

• Orders matched by best price
• Same price: first order wins
• Most common in equity markets

2. Pro-Rata

• Orders at same price get proportional fills
• Common in futures markets

3. Price-Time with Size Priority

• Larger orders get priority
• Encourages liquidity provision

Low-Latency Techniques

1. Kernel Bypass

• Standard network stack too slow
• Direct NIC access via DPDK or custom drivers
• Saves microseconds per message

2. Lock-Free Data Structures

• Locks cause latency spikes
• Lock-free queues and order books
• Careful memory ordering

3. Memory-Mapped Files

• Avoid system calls for persistence
• Direct memory access
• Operating system handles flushing

4. CPU Pinning

• Critical threads on dedicated cores
• Avoid context switching
• Disable hyperthreading on these cores

5. FPGA Acceleration

• Hardware-accelerated order processing
• Nanosecond latency for hot paths
• Common at top exchanges

Network Architecture

Multicast Market Data:

• One-to-many distribution
• UDP multicast for efficiency
• Sequence numbers for gap detection
• Redundant feeds for reliability

Co-location:

• Trading firms place servers in exchange data centers
• Minimize physical distance
• Equal cable lengths for fairness
• Measured in feet, not miles

Network Hardware:

• Cut-through switches (forward before fully received)
• 10/25/100 Gbps links
• Optimized network topology

Order Types

Market Order:

• Execute immediately at best available price
• Guaranteed execution, not price

Limit Order:

• Execute only at specified price or better
• May not fill if price never reached

Stop Order:

• Becomes market order when trigger price hit
• Used for risk management

Complex Orders:

• Iceberg (hidden size)
• IOC (immediate or cancel)
• FOK (fill or kill)
• Pegged orders

Fault Tolerance

High Availability Requirements:

• 99.999% uptime (5.26 minutes/year downtime)
• Seamless failover
• No lost orders or trades

Strategies:

1. Hot Standby

• Synchronized backup matching engine
• Sub-second failover

2. Order Replay

• All orders journaled
• Can rebuild state from journal
• Used for disaster recovery

3. Circuit Breakers

• Halt trading on extreme moves
• Prevent cascade failures
• Required by regulators

Regulatory Considerations

Audit Trail:

• Every order and trade logged
• Nanosecond timestamps
• Retained for years
• Available for regulatory review

Market Surveillance:

• Real-time manipulation detection
• Pattern analysis
• Insider trading detection

Best Execution:

• Obligation to get best price for customers
• Requires robust market data
• Documented execution quality

Scaling Challenges

Peak Processing:

• Opening/closing auctions
• Major news events
• Market volatility

Solutions:

• Over-provision capacity (10x normal)
• Graceful degradation
• Rate limiting (throttling)
• Queue management

Interview Application

When designing trading systems:

Key Requirements:

• Latency (microseconds)
• Throughput (millions of messages/second)
• Correctness (no lost orders, accurate matching)
• Availability (five 9s)
• Auditability (complete history)

Common Questions:

• Design a matching engine
• How to handle market data distribution
• Implementing price-time priority
• Handling system failures

Trade-offs:

• Latency vs throughput
• Consistency vs availability
• Feature richness vs simplicity

Stock exchange architecture represents the extreme of low-latency, high-reliability system design - techniques here apply (in diluted form) to many other systems.