如何加速 Pandas 多级数据帧总和？

Question

I am trying to speed up the sum for several big multilevel dataframes.

Here is a sample:

df1 = mul_df(5000,30,400) # mul_df to create a big multilevel dataframe
#let df2, df3, df4 = df1, df1, df1 to minimize the memory usage, 
#they can also be mul_df(5000,30,400) 
df2, df3, df4 = df1, df1, df1

In [12]: timeit df1+df2+df3+df4
1 loops, best of 3: 993 ms per loop

I am not satisfy with the 993ms, Is there any way to speed up ? Can cython improve the performance ? If yes, how to write the cython code ? Thanks.

Note : mul_df() is the function to create the demo multilevel dataframe.

import itertools
import numpy as np
import pandas as pd

def mul_df(level1_rownum, level2_rownum, col_num, data_ty='float32'):
    ''' create multilevel dataframe, for example: mul_df(4,2,6)'''

    index_name = ['STK_ID','RPT_Date']
    col_name = ['COL'+str(x).zfill(3) for x in range(col_num)]

    first_level_dt = [['A'+str(x).zfill(4)]*level2_rownum for x in range(level1_rownum)]
    first_level_dt = list(itertools.chain(*first_level_dt)) #flatten the list
    second_level_dt = ['B'+str(x).zfill(3) for x in range(level2_rownum)]*level1_rownum

    dt = pd.DataFrame(np.random.randn(level1_rownum*level2_rownum, col_num), columns=col_name, dtype = data_ty)
    dt[index_name[0]] = first_level_dt
    dt[index_name[1]] = second_level_dt

    rst = dt.set_index(index_name, drop=True, inplace=False)
    return rst

Update:

Data on my Pentium Dual-Core T4200@2.00GHZ, 3.00GB RAM, WindowXP, Python 2.7.4, Numpy 1.7.1, Pandas 0.11.0, numexpr 2.0.1 (Anaconda 1.5.0 (32-bit))

In [1]: from pandas.core import expressions as expr
In [2]: import numexpr as ne

In [3]: df1 = mul_df(5000,30,400)
In [4]: df2, df3, df4 = df1, df1, df1

In [5]: expr.set_use_numexpr(False)
In [6]: %timeit df1+df2+df3+df4
1 loops, best of 3: 1.06 s per loop

In [7]: expr.set_use_numexpr(True)
In [8]: %timeit df1+df2+df3+df4
1 loops, best of 3: 986 ms per loop

In [9]: %timeit  DataFrame(ne.evaluate('df1+df2+df3+df4'),columns=df1.columns,index=df1.index,dtype='float32')
1 loops, best of 3: 388 ms per loop

Answer 1

method 1: On my machine not so bad (with numexpr disabled)

In [41]: from pandas.core import expressions as expr

In [42]: expr.set_use_numexpr(False)

In [43]: %timeit df1+df2+df3+df4
1 loops, best of 3: 349 ms per loop

method 2: Using numexpr (which is by default enabled if numexpr is installed)

In [44]: expr.set_use_numexpr(True)

In [45]: %timeit df1+df2+df3+df4
10 loops, best of 3: 173 ms per loop

method 3: Direct use of numexpr

In [34]: import numexpr as ne

In [46]: %timeit  DataFrame(ne.evaluate('df1+df2+df3+df4'),columns=df1.columns,index=df1.index,dtype='float32')
10 loops, best of 3: 47.7 ms per loop

These speedups are achieved using numexpr because:

• avoids using intermediate temporary arrays (which in the case you are presenting is probably quite inefficient in numpy, I suspect this is being evaluated like ((df1+df2)+df3)+df4
• uses multi-cores as available

As I hinted above, pandas uses numexpr under the hood for certain types of ops (in 0.11), eg df1 + df2 would be evaluated this way, however the example you are giving here will result in several calls to numexpr (this is method 2 is faster than method 1.). Using the direct (method 3) ne.evaluate(...) achieves even more speedups.

Note that in pandas 0.13 (0.12 will be released this week), we are implemented a function pd.eval which will in effect do exactly what my example above does. Stay tuned (if you are adventurous this will be in master somewhat soon: https://github.com/pydata/pandas/pull/4037 )

In [5]: %timeit pd.eval('df1+df2+df3+df4')
10 loops, best of 3: 50.9 ms per loop

Lastly to answer your question, cython will not help here at all; numexpr is quite efficient at this type of problem (that said, there are situation where cython is helpful)

One caveat: in order to use the direct Numexpr method the frames should be already aligned (Numexpr operates on the numpy array and doesn't know anything about the indices). also they should be a single dtype

Answer 2

Other Observations

• You cannot expect more speedup if you have only 2 cores on your machine. In the end numexpression relies on parallelisation and the performant usage of the cpu cache.
• What you do is to some extend wrong. Numexpressions on DataFrames are fast, but wrong. They do not return the right result, if the DataFrames are not equally indexed. Different sorting will already trouble you, which I show below.
• If you add DataFrames with different indexes, the whole stuff is not that performant anymore. Well, Pandas does quite a good job to add the proper rows for you by looking up the corrsponding index entries. This comes with the natural cost.

In the following my Observations: - First, I reproduce your test case and come to other results. Using numexpression under the hood of Pandas increases performance significantly. - Second, I sort one of the four DataFrames in descending order and rerun all cases. The performance breaks, and additionally, (as expected) numexpression evaluation on Pandas DataFrames leads to wrong results.

Equal Indices on all Frames

This case reproduces your case. The only difference is, that I create copies of the inital DataFrame instance. So there is nothing shared. There are different objects (ids) in use to make sure, that numexpression can deal with it.

import itertools
import numpy as np
import pandas as pd

def mul_df(level1_rownum, level2_rownum, col_num, data_ty='float32'):
    ''' create multilevel dataframe, for example: mul_df(4,2,6)'''

    index_name = ['STK_ID','RPT_Date']
    col_name = ['COL'+str(x).zfill(3) for x in range(col_num)]

    first_level_dt = [['A'+str(x).zfill(4)]*level2_rownum for x in range(level1_rownum)]
    first_level_dt = list(itertools.chain(*first_level_dt)) #flatten the list
    second_level_dt = ['B'+str(x).zfill(3) for x in range(level2_rownum)]*level1_rownum

    dt = pd.DataFrame(np.random.randn(level1_rownum*level2_rownum, col_num), columns=col_name, dtype = data_ty)
    dt[index_name[0]] = first_level_dt
    dt[index_name[1]] = second_level_dt

    rst = dt.set_index(index_name, drop=True, inplace=False)
    return rst

df1 = mul_df(5000,30,400)
df2, df3, df4 = df1.copy(), df1.copy(), df1.copy() 

pd.options.compute.use_numexpr = False

%%timeit
df1 + df2 + df3 + df4

564 ms ± 10.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

pd.options.compute.use_numexpr = True

%%timeit 
df1 + df2 + df3 + df4

152 ms ± 1.47 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

import numexpr as ne

%%timeit
pd.DataFrame(ne.evaluate('df1 + df2 + df3 + df4'), columns=df1.columns, index=df1.index, dtype='float32')

66.4 ms ± 1.16 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

(df1 + df2 + df3 + df4).equals(pd.DataFrame(ne.evaluate('df1 + df2 + df3 + df4'), columns=df1.columns, index=df1.index, dtype='float32'))

True

(Slightly) Different Indices on some Frames

Here I sort one of the DataFrames in descending order, therefore changing the index and reshuffling the rows in the dataframe internal numpy array.

import itertools
import numpy as np
import pandas as pd

def mul_df(level1_rownum, level2_rownum, col_num, data_ty='float32'):
    ''' create multilevel dataframe, for example: mul_df(4,2,6)'''

    index_name = ['STK_ID','RPT_Date']
    col_name = ['COL'+str(x).zfill(3) for x in range(col_num)]

    first_level_dt = [['A'+str(x).zfill(4)]*level2_rownum for x in range(level1_rownum)]
    first_level_dt = list(itertools.chain(*first_level_dt)) #flatten the list
    second_level_dt = ['B'+str(x).zfill(3) for x in range(level2_rownum)]*level1_rownum

    dt = pd.DataFrame(np.random.randn(level1_rownum*level2_rownum, col_num), columns=col_name, dtype = data_ty)
    dt[index_name[0]] = first_level_dt
    dt[index_name[1]] = second_level_dt

    rst = dt.set_index(index_name, drop=True, inplace=False)
    return rst

df1 = mul_df(5000,30,400)
df2, df3, df4 = df1.copy(), df1.copy(), df1.copy().sort_index(ascending=False)

pd.options.compute.use_numexpr = False

%%timeit
df1 + df2 + df3 + df4

1.36 s ± 67.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

pd.options.compute.use_numexpr = True

%%timeit 
df1 + df2 + df3 + df4

928 ms ± 39.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

import numexpr as ne

%%timeit
pd.DataFrame(ne.evaluate('df1 + df2 + df3 + df4'), columns=df1.columns, index=df1.index, dtype='float32')

68 ms ± 2.2 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

(df1 + df2 + df3 + df4).equals(pd.DataFrame(ne.evaluate('df1 + df2 + df3 + df4'), columns=df1.columns, index=df1.index, dtype='float32'))

False

Conclusions

By using numexpr

• Quite some speedup is gained when operating on equally indexed DataFrames.
• The same is true if you have other expressions with a single dataframe, as 2 * df1 .
• This is not the case if operations between DataFrames with different indices are used.
• It leads even to completely wrong results if one evaluates expressions containing Pandas DataFrames. By chance they can be right. But numexpression is made for optimizing expressions on Numpy arrays.