Colocation and Hardware Acceleration

In high-frequency trading, the physical distance between your server and the exchange's matching engine matters. A millisecond is a mile. Firms spend millions colocating to shave microseconds.

Colocation

Colocation is the practice of placing your trading servers in the same data center as the exchange. The closer your server is to the exchange's matching engine, the lower your network latency.

Physical Distance

Light travels at approximately 200,000 km/s in fiber optic cables (slower than vacuum due to the refractive index).

NY to Chicago: ~1,100 km → min latency ≈ 7.5 ms
NY to London: ~5,500 km → min latency ≈ 28 ms

Colocation can reduce this to single-digit microseconds within the same data center.

Rack Placement

Even within a data center, rack position matters:

• Front-rack placement minimizes cable length to the exchange switch
• Same row as the exchange's matching engine servers
• Cross-connects — direct fiber cables between cages

Kernel Bypass

The Linux kernel adds latency to network I/O: context switches, buffer copies, interrupt handling. Kernel bypass eliminates this:

• DPDK (Data Plane Development Kit) — userspace networking
• Solarflare Onload — TCP stack in userspace
• Solarflare OpenOnload — kernel-bypass library

Traditional: NIC → Kernel → App → Kernel → NIC  (~20-50 μs)
Kernel bypass: NIC → App → NIC  (~1-5 μs)

FPGA Acceleration

Field-Programmable Gate Arrays (FPGAs) implement logic in hardware — orders of magnitude faster than CPUs for specific tasks:

// CPU: parse market data in software → microseconds
// FPGA: parse market data in hardware → nanoseconds
 
FPGA use cases:
1. Market data parsing → nanosecond order book updates
2. Risk checks → nanoseconds per order
3. Ping-pong strategies → microseconds round-trip

Microwave Towers

Between major financial centers, some firms deploy microwave towers — straight-line paths that are shorter than fiber routes:

NY to Chicago:
  Fiber: ~1,400 km → ~7.5 ms
  Microwave: ~1,100 km → ~6.0 ms
  Time saved: 1.5 ms

That 1.5ms translates to millions in annual profits for a high-volume strategy.

The Latency Stack

Every microsecond is fought for. The fastest firms deploy FPGAs, custom hardware, and microwave towers. But the gains follow diminishing returns — the last microsecond costs more than the first hundred.