[Bench] Reference for perfomance analysis

This is an example that compares HDK with Pandas performance.
Can be modified by increasing N (originally was 10).

See also: #567

Signed-off-by: Dmitrii Makarenko <[email protected]>

Author: Devjiu

join_compileWU_slow_ref.py

@@ -0,0 +1,130 @@
+#!/usr/bin/env python
+
+import sys
+import pyhdk
+import pandas as pd
+import numpy as np
+import time
+
+
+def compare_tables(left_df: pd.DataFrame, right_df: pd.DataFrame):
+    try_to_guess = True
+    left_cols = left_df.columns.to_list()
+    right_cols = right_df.columns.to_list()
+    left_cols.sort()
+    right_cols.sort()
+
+    diff_idx = [
+        idx for idx, col_name in enumerate(right_cols) if col_name != left_cols[idx]
+    ]
+
+    print("compare lists: ", diff_idx)
+    drop_left = []
+    drop_right = []
+    for drop_idx in diff_idx:
+        drop_left += [left_cols[drop_idx]]
+        drop_right += [right_cols[drop_idx]]
+    if try_to_guess:
+        right_df = right_df.rename(columns=dict(zip(drop_right, drop_left)))
+    else:
+        print("cols: ", left_cols, " drops: ", drop_left)
+        print("cols: ", right_cols, " drops: ", drop_right)
+        left_df = left_df.drop(columns=drop_left)
+        right_df = right_df.drop(columns=drop_right)
+
+    left_cols = left_df.columns.to_list()
+    right_cols = right_df.columns.to_list()
+    left_cols.sort()
+    right_cols.sort()
+
+    print("cols: r - ", right_cols, " l - ", left_cols)
+
+    assert left_cols == right_cols, "Table column names are different"
+
+    left_df.sort_values(by=left_cols, inplace=True)
+    right_df.sort_values(by=left_cols, inplace=True)
+    for col in left_df.columns:
+        if left_df[col].dtype in ["category"]:
+            left_df[col] = left_df[col].astype("str")
+            right_df[col] = right_df[col].astype("str")
+    print("l dtypes \n", left_df.dtypes)
+    print("r dtypes \n", right_df.dtypes)
+
+    print("l size: ", left_df.size, " - r size: ", right_df.size)
+
+    left_df = left_df.reset_index(drop=True)
+    right_df = right_df.reset_index(drop=True)
+    if not all(left_df == right_df):
+        mask = left_df == right_df
+        print("Mismathed left: ")
+        print(left_df[mask])
+        print("         right: ")
+        print(left_df[mask])
+        raise RuntimeError("Results mismatched")
+
+
+pyhdk_init_args = {}
+pyhdk_init_args["enable_debug_timer"] = True
+pyhdk_init_args["enable_cpu_groupby_multifrag_kernels"] = False
+# pyhdk_init_args["debug_logs"] = True
+hdk = pyhdk.init(**pyhdk_init_args)
+fragment_size = 4000000
+
+N = 2
+
+np.random.seed(1)
+column_list = list()
+for num in range(N):
+    df_setup = {
+        "column_1": np.random.randint(0, 150, size=(15000)),
+        "column_3": np.random.randint(0, 6, size=(15000)),
+        "column_5": np.random.randint(0, 10, size=(15000)),
+        "A" + str(num): np.random.randint(0, 100, size=(15000)),
+        "B" + str(num): np.random.randint(0, 100, size=(15000)),
+        "C" + str(num): np.random.randint(0, 100, size=(15000)),
+        "D" + str(num): np.random.randint(0, 100, size=(15000)),
+    }
+    column_list.append(df_setup)
+
+df_list = list()
+for num in range(N):
+    df = pd.DataFrame(column_list[num])
+    df_list.append(df)
+
+t1 = time.time()
+for idx, df in enumerate(df_list):
+    if idx == 0:
+        df_base = df.copy()
+        df_base = df_base.to_dict("list")
+    else:
+        if type(df_base) == dict:
+            ht_base = hdk.import_pydict(df_base)
+        else:
+            ht_base = df_base
+        df_r = hdk.import_pydict(df.to_dict("list"))
+        df_ans = ht_base.join(
+            df_r,
+            ["column_1", "column_3", "column_5"],
+            ["column_1", "column_3", "column_5"],
+        ).run()
+        df_base = df_ans.to_arrow().to_pandas().to_dict("list")
+print("Hdk Time:", time.time() - t1)
+
+t2 = time.time()
+for idx, df in enumerate(df_list):
+    if idx == 0:
+        df_base_pd = df.copy()
+    else:
+        df_base_pd = pd.merge(
+            df_base_pd,
+            df,
+            left_on=["column_1", "column_3", "column_5"],
+            right_on=["column_1", "column_3", "column_5"],
+            how="inner",
+        )
+print("Pandas Time:", time.time() - t2)
+
+print("[hdk] shape: ", pd.DataFrame(df_base).shape)
+print("[ pd] shape: ", df_base_pd.shape)
+print("compare: ")
+compare_tables(pd.DataFrame(df_base), df_base_pd)