性能追求第二部分:Perl 与 Python
运行了一个玩具性能示例后,我们现在将稍微偏离主题并将性能与
进行对比一些 python 实现。首先让我们设置计算阶段,并提供命令行
python 脚本的功能。
import argparseimport timeimport mathimport numpy as npimport osfrom numba import njitfrom joblib import parallel, delayedparser = argparse.argumentparser()parser.add_argument("--workers", type=int, default=8)parser.add_argument("--arraysize", type=int, default=100_000_000)args = parser.parse_args()# set the number of threads to 1 for different librariesprint("=" * 80)print( f"starting the benchmark for {args.arraysize} elements " f"using {args.workers} threads/workers")# generate the data structures for the benchmarkarray0 = [np.random.rand() for _ in range(args.arraysize)]array1 = array0.copy()array2 = array0.copy()array_in_np = np.array(array1)array_in_np_copy = array_in_np.copy()
这是我们的参赛者:
for i in range(len(array0)): array0[i] = math.cos(math.sin(math.sqrt(array0[i])))
np.sqrt(array_in_np, out=array_in_np)np.sin(array_in_np, out=array_in_np)np.cos(array_in_np, out=array_in_np)
def compute_inplace_with_joblib(chunk): return np.cos(np.sin(np.sqrt(chunk))) #parallel function for joblibchunks = np.array_split(array1, args.workers) # split the array into chunksnumresults = parallel(n_jobs=args.workers)( delayed(compute_inplace_with_joblib)(chunk) for chunk in chunks )# process each chunk in a separate threadarray1 = np.concatenate(numresults) # concatenate the results
@njitdef compute_inplace_with_numba(array): np.sqrt(array,array) np.sin(array,array) np.cos(array,array) ## njit will compile this function to machine codecompute_inplace_with_numba(array_in_np_copy)
这是计时结果:
in place in ( base python): 11.42 secondsin place in (python joblib): 4.59 secondsin place in ( python numba): 2.62 secondsin place in ( python numpy): 0.92 seconds
numba 出奇的慢!?难道是由于 mohawk2 在 irc 交流中关于此问题指出的编译开销造成的吗?
为了测试这一点,我们应该在执行基准测试之前调用compute_inplace_with_numba一次。这样做表明 numba 现在比 numpy 更快。
in place in ( base python): 11.89 secondsin place in (python joblib): 4.42 secondsin place in ( python numpy): 0.93 secondsin place in ( python numba): 0.49 seconds
n<-50000000x<-runif(n)start_time <- sys.time()result <- cos(sin(sqrt(x)))end_time <- sys.time()# calculate the time takentime_taken <- end_time - start_time# print the time takenprint(sprintf("time in base r: %.2f seconds", time_taken))
Time in base R: 1.30 seconds