results: algoritm comparison plots and plotting scripts
This commit is contained in:
Erik Fabrizzi
104
results-and-plotting/python/scripts/plot_alg.py
Normal file
104
results-and-plotting/python/scripts/plot_alg.py
Normal file
@@ -0,0 +1,104 @@
|
||||
from matplotlib.ticker import FuncFormatter
|
||||
import pandas as pd
|
||||
import matplotlib.pyplot as plt
|
||||
import numpy as np
|
||||
from matplotlib.lines import Line2D
|
||||
from enum import Enum
|
||||
|
||||
|
||||
class columns ():
|
||||
benchmark_type = 'benchmark_type'
|
||||
proc_num = 'proc_num'
|
||||
msg_size_bytes = 'msg_size_bytes'
|
||||
repetitions = 'repetitions'
|
||||
t_min_usec = 't_min_usec'
|
||||
t_max_usec = 't_max_usec'
|
||||
t_avg_usec = 't_avg_usec'
|
||||
mpi_datatype = 'mpi_datatype'
|
||||
mpi_red_datatype = 'mpi_red_datatype'
|
||||
mpi_red_op = 'mpi_red_op'
|
||||
creation_time = 'creation_time'
|
||||
n_nodes = 'n_nodes'
|
||||
off_cache_flag = 'off_cache_flag'
|
||||
algorithm = 'algorithm'
|
||||
|
||||
|
||||
class collectives(Enum):
|
||||
Bcast = 'Bcast'
|
||||
Reduce = 'Reduce'
|
||||
Allreduce = 'Allreduce'
|
||||
Alltoall = 'Alltoall'
|
||||
Scatter = 'Scatter'
|
||||
Reduce_scatter = 'Reduce_scatter'
|
||||
Allgather = 'Allgather'
|
||||
Gather = 'Gather'
|
||||
|
||||
|
||||
data_file = "./data/data_04_11_25_algs.csv"
|
||||
|
||||
df_multinode = pd.read_csv(data_file, delimiter=',')
|
||||
df_multinode.fillna(0, inplace=True)
|
||||
df_multinode = df_multinode[df_multinode[columns.off_cache_flag] == 50]
|
||||
for c in collectives:
|
||||
|
||||
df_single = df_multinode[df_multinode[columns.benchmark_type]
|
||||
== c.value]
|
||||
df_single = df_single[df_single[columns.msg_size_bytes] > 1000]
|
||||
df_single = df_single[[columns.proc_num, columns.msg_size_bytes,
|
||||
columns.t_avg_usec, columns.algorithm]]
|
||||
|
||||
df_gather_best = df_single.loc[ # pyright: ignore
|
||||
df_single.groupby([columns.msg_size_bytes, columns.proc_num])[ # pyright: ignore
|
||||
columns.t_avg_usec].idxmin()
|
||||
].reset_index(drop=True)
|
||||
|
||||
df_gather_best = df_gather_best[df_gather_best[columns.msg_size_bytes] > 2**17]
|
||||
|
||||
pivot_best = df_gather_best.pivot(index=columns.msg_size_bytes, # pyright: ignore
|
||||
columns=columns.proc_num, values=columns.t_avg_usec)
|
||||
|
||||
X = pivot_best.columns.values # proc_num
|
||||
Y = pivot_best.index.values # msg_size_bytes
|
||||
X, Y = np.meshgrid(X, Y) # pyright: ignore
|
||||
Z = pivot_best.values
|
||||
|
||||
alg_pivot = df_gather_best.pivot(
|
||||
index=columns.msg_size_bytes,
|
||||
columns=columns.proc_num,
|
||||
values=columns.algorithm
|
||||
)
|
||||
|
||||
algorithms = alg_pivot.values.flatten()
|
||||
unique_algs = sorted(pd.unique(algorithms[~pd.isna(algorithms)]))
|
||||
color_map = {alg: i for i, alg in enumerate(unique_algs)}
|
||||
color_values = np.array([color_map.get(a, np.nan) for a in algorithms])
|
||||
|
||||
fig = plt.figure(figsize=(16, 9))
|
||||
ax = fig.add_subplot(111, projection='3d')
|
||||
surf = ax.plot_wireframe(X, Y, Z,
|
||||
color='black', linewidths=1)
|
||||
surf_points = ax.scatter(X,
|
||||
Y, Z, c=color_values, cmap='viridis', s=20, depthshade=False) # pyright: ignore
|
||||
|
||||
handles = [
|
||||
Line2D([0], [0],
|
||||
marker='o', color='w',
|
||||
label=alg,
|
||||
markerfacecolor=plt.cm.viridis(
|
||||
color_map[alg] / max(len(unique_algs)-1, 1)),
|
||||
markersize=8)
|
||||
for alg in unique_algs
|
||||
]
|
||||
|
||||
ax.legend(handles=handles, title="Algorithm", loc='upper right')
|
||||
ax.set_xlabel("Process Count")
|
||||
ax.set_ylabel("Message Size [B]")
|
||||
ax.set_zlabel("Average Time [μs]")
|
||||
ax.set_title(f"{c.value}")
|
||||
ax.set_xticks(pivot_best.columns.values) # pyright: ignore
|
||||
ax.set_xticklabels(pivot_best.columns.values)
|
||||
ax.set_yticks(Y[:, 0])
|
||||
ymin, ymax = ax.get_ylim()
|
||||
ax.set_ylim(ymin*0.8, ymax) # 30% more space at top
|
||||
ax.set_yticklabels([f"$2^{{{int(np.log2(v))}}}$" for v in Y[:, 0]])
|
||||
plt.savefig(f"./plots/{c.value.lower()}_algcomp.png")
|
||||
127
results-and-plotting/python/scripts/plot_alg_log.py
Normal file
127
results-and-plotting/python/scripts/plot_alg_log.py
Normal file
@@ -0,0 +1,127 @@
|
||||
from matplotlib.ticker import FuncFormatter
|
||||
import pandas as pd
|
||||
import matplotlib.pyplot as plt
|
||||
import numpy as np
|
||||
from matplotlib.lines import Line2D
|
||||
from enum import Enum
|
||||
|
||||
|
||||
class columns ():
|
||||
benchmark_type = 'benchmark_type'
|
||||
proc_num = 'proc_num'
|
||||
msg_size_bytes = 'msg_size_bytes'
|
||||
repetitions = 'repetitions'
|
||||
t_min_usec = 't_min_usec'
|
||||
t_max_usec = 't_max_usec'
|
||||
t_avg_usec = 't_avg_usec'
|
||||
mpi_datatype = 'mpi_datatype'
|
||||
mpi_red_datatype = 'mpi_red_datatype'
|
||||
mpi_red_op = 'mpi_red_op'
|
||||
creation_time = 'creation_time'
|
||||
n_nodes = 'n_nodes'
|
||||
off_cache_flag = 'off_cache_flag'
|
||||
algorithm = 'algorithm'
|
||||
|
||||
|
||||
class collectives(Enum):
|
||||
Bcast = 'Bcast'
|
||||
Reduce = 'Reduce'
|
||||
Allreduce = 'Allreduce'
|
||||
Alltoall = 'Alltoall'
|
||||
Scatter = 'Scatter'
|
||||
Reduce_scatter = 'Reduce_scatter'
|
||||
Allgather = 'Allgather'
|
||||
Gather = 'Gather'
|
||||
|
||||
|
||||
def log_notation(val, pos):
|
||||
return f" $1e{int(val)}$" if val != 0 else "1"
|
||||
|
||||
|
||||
def log2_notation(val, pos):
|
||||
return "$2^{"+str(int(val))+"}$" if val != 0 else "1"
|
||||
|
||||
|
||||
data_file = "./data/data_04_11_25_algs.csv"
|
||||
|
||||
df_multinode = pd.read_csv(data_file, delimiter=',')
|
||||
df_multinode.fillna(0, inplace=True)
|
||||
df_multinode = df_multinode[df_multinode[columns.off_cache_flag] == 50]
|
||||
for c in collectives:
|
||||
|
||||
df_single = df_multinode[df_multinode[columns.benchmark_type]
|
||||
== c.value]
|
||||
df_single = df_single[df_single[columns.msg_size_bytes] > 1000]
|
||||
df_single = df_single[[columns.proc_num, columns.msg_size_bytes,
|
||||
columns.t_avg_usec, columns.algorithm]]
|
||||
|
||||
df_gather_best = df_single.loc[ # pyright: ignore
|
||||
df_single.groupby([columns.msg_size_bytes, columns.proc_num])[ # pyright: ignore
|
||||
columns.t_avg_usec].idxmin()
|
||||
].reset_index(drop=True)
|
||||
|
||||
df_gather_worst = df_single.loc[ # pyright: ignore
|
||||
df_single.groupby([columns.msg_size_bytes, columns.proc_num])[ # pyright: ignore
|
||||
columns.t_avg_usec].idxmax()
|
||||
].reset_index(drop=True)
|
||||
# df_gather_select = df_gather_select[df_gather_select[columns.msg_size_bytes] > 2**17]
|
||||
|
||||
pivot_best = df_gather_best.pivot(index=columns.msg_size_bytes, # pyright: ignore
|
||||
columns=columns.proc_num, values=columns.t_avg_usec)
|
||||
pivot_worst = df_gather_worst.pivot(index=columns.msg_size_bytes, # pyright: ignore
|
||||
columns=columns.proc_num, values=columns.t_avg_usec)
|
||||
|
||||
X = pivot_best.columns.values # proc_num
|
||||
Y = pivot_best.index.values # msg_size_bytes
|
||||
X, Y = np.meshgrid(X, Y) # pyright: ignore
|
||||
Z = pivot_best.values
|
||||
|
||||
X_w = pivot_worst.columns.values # proc_num
|
||||
Y_w = pivot_worst.index.values # msg_size_bytes
|
||||
X_w, Y_w = np.meshgrid(X_w, Y_w) # pyright: ignore
|
||||
Z_w = pivot_worst.values
|
||||
|
||||
alg_pivot = df_gather_best.pivot(
|
||||
index=columns.msg_size_bytes,
|
||||
columns=columns.proc_num,
|
||||
values=columns.algorithm
|
||||
)
|
||||
|
||||
algorithms = alg_pivot.values.flatten()
|
||||
unique_algs = sorted(pd.unique(algorithms[~pd.isna(algorithms)]))
|
||||
color_map = {alg: i for i, alg in enumerate(unique_algs)}
|
||||
color_values = np.array([color_map.get(a, np.nan) for a in algorithms])
|
||||
|
||||
fig = plt.figure(figsize=(16, 9))
|
||||
ax = fig.add_subplot(111, projection='3d')
|
||||
surf = ax.plot_wireframe(X, np.log2(Y), np.log(Z),
|
||||
color='black', linewidths=1)
|
||||
surf = ax.plot_wireframe(X_w, np.log2(
|
||||
Y_w), np.log(Z_w), color='gray', linewidths=0.3)
|
||||
|
||||
surf_points = ax.scatter(X, np.log2(
|
||||
Y), np.log(Z), c=color_values, cmap='viridis', s=20, depthshade=False) # pyright: ignore
|
||||
|
||||
surf_points = ax.scatter(X_w, np.log2(
|
||||
Y_w), np.log(Z_w), c='gray', alpha=0.2, s=20, depthshade=False) # pyright: ignore
|
||||
|
||||
handles = [
|
||||
Line2D([0], [0],
|
||||
marker='o', color='w',
|
||||
label=alg,
|
||||
markerfacecolor=plt.cm.viridis(
|
||||
color_map[alg] / max(len(unique_algs)-1, 1)),
|
||||
markersize=8)
|
||||
for alg in unique_algs
|
||||
]
|
||||
|
||||
ax.legend(handles=handles, title="Algorithm", loc='upper right')
|
||||
ax.set_xlabel("Process Count")
|
||||
ax.set_ylabel("Message Size [B] (log2)")
|
||||
ax.set_zlabel("Average Time [μs] (log10)")
|
||||
ax.set_title(f"{c.value}")
|
||||
ax.set_xticks(pivot_best.columns.values) # pyright: ignore
|
||||
ax.set_xticklabels(pivot_best.columns.values)
|
||||
ax.yaxis.set_major_formatter(FuncFormatter(log2_notation))
|
||||
ax.zaxis.set_major_formatter(FuncFormatter(log_notation))
|
||||
plt.savefig(f"./plots/{c.value.lower()}_algcomp_log.png")
|
||||
Reference in New Issue
Block a user