Skip to content

Latest commit

 

History

History
94 lines (68 loc) · 2.6 KB

README.md

File metadata and controls

94 lines (68 loc) · 2.6 KB

Zero-Copy Market Data Processing

This component implements high-performance, low-latency market data processing using CUDA's zero-copy memory capabilities and asynchronous streams. It significantly reduces latency for high-frequency trading (HFT) systems.

Technical Overview

Key Features

  1. Zero-Copy Memory Access

    • Uses mapped memory for direct GPU access to host memory
    • Eliminates costly PCIe bus transfers
    • Reduces latency by removing memory copy overhead

  2. Pipelined Processing with CUDA Streams

    • Implements asynchronous data processing using multiple streams
    • Overlaps computation and data transfer operations
    • Enables continuous processing of market data feeds

  3. Asynchronous API

    • Thread-safe queue for submitting market data packets
    • Background processing thread for continuous operation
    • Non-blocking design for real-time market data handling

Performance Goal

  • Reduce end-to-end latency by 50% compared to standard memory transfer approaches

Implementation Details

Memory Management Techniques

Standard Memory Transfers (Baseline)

CPU Memory -> cudaMemcpy -> GPU Memory -> Processing -> cudaMemcpy -> CPU Memory

Zero-Copy Approach

Mapped Memory (accessible by both CPU and GPU) -> In-place Processing

CUDA Streams for Pipelining

The implementation uses multiple CUDA streams to enable overlapped execution of:

  • Memory operations (when needed)
  • Kernel execution for different data segments
  • Event synchronization

Market Data Format

The system processes a simplified ITCH-like market data format with these fields:

  • Timestamp (nanosecond precision)
  • Message type (Add, Cancel, Execute, Trade)
  • Order ID
  • Side (Buy/Sell)
  • Price
  • Quantity

Benchmark Results

Typical results for 1,000,000 market data packets:

Method Average Latency (μs) Improvement
Standard ~13,000 μs Baseline
Zero-Copy ~3,700 μs ~69%
Zero-Copy + Streams ~3,800 μs ~69%

Building and Running

Prerequisites

  • CUDA Toolkit 11.0+
  • C++17 compatible compiler
  • CMake 3.18+

Compilation

mkdir build && cd build
cmake ..
cmake --build .

Running the Benchmark

./zero_copy_processor [num_packets] [num_runs]

Integration with Order Book

This zero-copy processor can be integrated with the existing order book implementation:

  1. Use zero-copy to efficiently receive and parse market data
  2. Feed processed orders directly into the GPU-accelerated order book
  3. Implement a continuous matching engine that processes order book updates