基准测试#

类似数据库操作的基准测试#

我们重现了类似数据库操作的基准测试，包括使用cudf.pandas的解决方案。以下是结果：

duckdb-benchmark-groupby-join

Results of the Database-like ops benchmark including cudf.pandas.

注意： 结果中某个解决方案缺少条形图表示在执行该解决方案的一个或多个查询时遇到了错误。

您可以在此处查看每个查询的结果 here.

重现步骤#

以下是重现cudf.pandas结果的步骤。重现其他解决方案结果的步骤记录在duckdblabs/db-benchmark中。

  1. 克隆最新的 duckdblabs/db-benchmark

  2. 为pandas构建环境：

virtualenv pandas/py-pandas

  1. 激活 pandas 虚拟环境：

source pandas/py-pandas/bin/activate

  1. 安装 cudf:

pip install --extra-index-url=https://pypi.nvidia.com cudf-cu12  # or cudf-cu11

  1. 修改pandas的join/group代码以使用cudf.pandas并移除dtype_backend关键字参数（不支持）：

diff --git a/pandas/groupby-pandas.py b/pandas/groupby-pandas.py
index 58eeb26..2ddb209 100755
--- a/pandas/groupby-pandas.py
+++ b/pandas/groupby-pandas.py
@@ -1,4 +1,4 @@
-#!/usr/bin/env python3
+#!/usr/bin/env -S python3 -m cudf.pandas

 print("# groupby-pandas.py", flush=True)

diff --git a/pandas/join-pandas.py b/pandas/join-pandas.py
index f39beb0..a9ad651 100755
--- a/pandas/join-pandas.py
+++ b/pandas/join-pandas.py
@@ -1,4 +1,4 @@
-#!/usr/bin/env python3
+#!/usr/bin/env -S python3 -m cudf.pandas

 print("# join-pandas.py", flush=True)

@@ -26,7 +26,7 @@ if len(src_jn_y) != 3:

 print("loading datasets " + data_name + ", " + y_data_name[0] + ", " + y_data_name[1] + ", " + y_data_name[2], flush=True)

-x = pd.read_csv(src_jn_x, engine='pyarrow', dtype_backend='pyarrow')
+x = pd.read_csv(src_jn_x, engine='pyarrow')

 # x['id1'] = x['id1'].astype('Int32')
 # x['id2'] = x['id2'].astype('Int32')
@@ -35,17 +35,17 @@ x['id4'] = x['id4'].astype('category') # remove after datatable#1691
 x['id5'] = x['id5'].astype('category')
 x['id6'] = x['id6'].astype('category')

-small = pd.read_csv(src_jn_y[0], engine='pyarrow', dtype_backend='pyarrow')
+small = pd.read_csv(src_jn_y[0], engine='pyarrow')
 # small['id1'] = small['id1'].astype('Int32')
 small['id4'] = small['id4'].astype('category')
 # small['v2'] = small['v2'].astype('float64')
-medium = pd.read_csv(src_jn_y[1], engine='pyarrow', dtype_backend='pyarrow')
+medium = pd.read_csv(src_jn_y[1], engine='pyarrow')
 # medium['id1'] = medium['id1'].astype('Int32')
 # medium['id2'] = medium['id2'].astype('Int32')
 medium['id4'] = medium['id4'].astype('category')
 medium['id5'] = medium['id5'].astype('category')
 # medium['v2'] = medium['v2'].astype('float64')
-big = pd.read_csv(src_jn_y[2], engine='pyarrow', dtype_backend='pyarrow')
+big = pd.read_csv(src_jn_y[2], engine='pyarrow')
 # big['id1'] = big['id1'].astype('Int32')
 # big['id2'] = big['id2'].astype('Int32')
 # big['id3'] = big['id3'].astype('Int32')

  1. 运行修改后的pandas基准测试：

./_launcher/solution.R --solution=pandas --task=groupby --nrow=1e7
./_launcher/solution.R --solution=pandas --task=groupby --nrow=1e8
./_launcher/solution.R --solution=pandas --task=join --nrow=1e7
./_launcher/solution.R --solution=pandas --task=join --nrow=1e8