为何Cython类属性访问慢于Python数据类？如何优化？

测试代码 #

Cython cdef类实现 #

cdef class BookL1:
    cdef readonly str exchange
    cdef readonly str symbol
    cdef readonly double bid
    cdef readonly double ask
    cdef readonly double bid_size
    cdef readonly double ask_size

    def __init__(self, str exchange, str symbol, double bid, double ask, double bid_size, double ask_size):
        self.exchange = exchange
        self.symbol = symbol
        self.bid = bid
        self.ask = ask
        self.bid_size = bid_size
        self.ask_size = ask_size

Python带slots的数据类实现 #

from dataclasses import dataclass

@dataclass(slots=True)
class BookL1:
    exchange: str
    symbol: str
    bid: float
    ask: float
    bid_size: float
    ask_size: float

性能测试代码 #

from trade_types import BookL1 as CBookL1
from utils import BookL1
import timeit

def create_book_l1():
    return BookL1("binance", "BTC/USDT", 30000.0, 30001.0, 1.5, 2.0)

def create_cbook_l1():
    return CBookL1("binance", "BTC/USDT", 30000.0, 30001.0, 1.5, 2.0)

def access_book_l1(book):
    return (book.exchange, book.symbol, book.bid, book.ask, book.bid_size, book.ask_size)

def run_benchmark(class_type, create_func, iterations=1000000):
    creation_time = timeit.timeit(create_func, number=iterations)
    
    instance = create_func()
    access_time = timeit.timeit(lambda: access_book_l1(instance), number=iterations)
    
    print(f"{class_type} Benchmark Results:")
    print(f"Creation time: {creation_time:.6f} seconds")
    print(f"Access time:   {access_time:.6f} seconds")
    print()

if __name__ == "__main__":
    run_benchmark("Python BookL1", create_book_l1)
    run_benchmark("Cython CBookL1", create_cbook_l1)

测试环境与结果 #

测试环境：Apple M3芯片 + macOS 14.6.1

测试结果：

Python BookL1 Benchmark Results:
Creation time: 0.200142 seconds
Access time:   0.105723 seconds

Cython CBookL1 Benchmark Results:
Creation time: 0.064651 seconds
Access time:   0.169590 seconds

原因分析 #

1. Python层属性访问开销：你定义的Cython类用了readonly修饰属性，这会生成Python风格的getter方法，每次访问属性都要走Python的属性查找流程——包括类型校验、方法调用等开销。而带slots的Python类，属性直接映射到固定内存位置，跳过了字典查找，访问路径更短。
2. 静态类型约束缺失：测试用的access_book_l1函数接收的是泛型Python对象，Cython没法在编译期确定传入的是Cython的BookL1实例，只能按Python对象处理属性访问，完全没用到Cython的静态优化能力。

优化方案 #

1. 用Cython静态类型函数处理属性访问 #

在Cython模块里直接写带静态类型的访问函数，让编译器直接生成内存访问代码，绕过Python层开销：

# 在trade_types.pyx中新增
cdef tuple _access_cbook_l1(BookL1 book):
    return (book.exchange, book.symbol, book.bid, book.ask, book.bid_size, book.ask_size)

# 暴露给Python调用
def access_cbook_l1(BookL1 book):
    return _access_cbook_l1(book)

然后修改测试代码，调用这个Cython实现的函数，替代原来的access_book_l1。

2. 在Cython类中添加批量属性获取方法 #

给Cython类加一个cdef方法，一次性返回所有属性，减少多次跨层访问的开销：

cdef class BookL1:
    # 原有属性和__init__不变
    
    cdef tuple get_attrs(self):
        return (self.exchange, self.symbol, self.bid, self.ask, self.bid_size, self.ask_size)
    
    # 可选：暴露给Python的接口
    def py_get_attrs(self):
        return self.get_attrs()

测试时直接调用get_attrs或py_get_attrs，一次拿到所有属性，避免多次单独访问的开销。

3. 优化Cython编译参数 #

在setup.py里添加激进的编译优化选项，让编译器生成更高效的机器码：

from setuptools import setup
from Cython.Build import cythonize

setup(
    ext_modules=cythonize(
        "trade_types.pyx",
        compiler_directives={"language_level": "3"},
        annotate=True
    ),
    extra_compile_args=["-O3", "-march=native"]
)

优化效果 #

做完以上优化后，Cython类的属性访问性能会明显提升——静态类型约束让Cython跳过了Python的属性查找流程，直接访问内存中的属性值，彻底消除跨层开销。

内容的提问来源于stack exchange，提问作者river7816