ErikFabrizzi/IMB-Benchmarking-tools
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Added data,plots and plotting scripts

Browse Source
This commit is contained in:
Erik Fabrizzi
2025-10-17 13:57:23 +02:00
parent a25f8ffec6
commit 459f12b086
4 changed files with 292 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
results-and-plotting/python/scripts/plot_all.py Normal file
View File

@@ -0,0 +1,69 @@
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from scipy.optimize import curve_fit
import matplotlib.cm as cm
def max_transfer_size(msg_size, np_procs, benchmark_type):
if benchmark_type == 'Allgather':
return (np_procs-72)*msg_size
elif benchmark_type == 'Scatter':
return (np_procs-72)*msg_size # ?
elif benchmark_type == 'Alltoall':
return 72*(np_procs-72)*msg_size
elif benchmark_type == 'Bcast':
return msg_size
elif benchmark_type == 'Gather':
return (np_procs)*msg_size # ?
elif benchmark_type == 'Reduce_scatter':
return 0.25*(np_procs-72)*(1/72)*msg_size # ?
elif benchmark_type == 'Allreduce':
return 0.25*(np_procs-72)*(1/72)*msg_size
elif benchmark_type == 'Reduce':
return 0.25*(np_procs-72)*(1/72)*msg_size
data_file = "data/data-multi-defand100cflag.csv"
df_multinode = pd.read_csv(data_file, delimiter=',')
df_multinode_offdef = df_multinode[df_multinode['off_cache_flag'] == 100]
benchmarks = df_multinode_offdef['benchmark_type'].unique().tolist()
benchmarks = [x for x in benchmarks if x[-1] != 'v']
print(benchmarks)
df_multinode_offdef = df_multinode_offdef[df_multinode_offdef['benchmark_type'].isin(
benchmarks)][df_multinode_offdef['msg_size_bytes'] > 1000]
df_multinode_offdef["max_transfer"] = df_multinode_offdef.apply(
lambda row: max_transfer_size(
msg_size=row["msg_size_bytes"],
np_procs=row["proc_num"],
benchmark_type=row["benchmark_type"]
),
axis=1
)
df_multinode_offdef["bytes/usec"] = df_multinode_offdef["max_transfer"] / \
df_multinode_offdef["t_avg_usec"]
df_multinode_offdef = df_multinode_offdef[df_multinode_offdef['benchmark_type']!='Allgather'][df_multinode_offdef['benchmark_type']!='Alltoall']
df_multinode_offdef = df_multinode_offdef[['benchmark_type','msg_size_bytes','t_avg_usec','proc_num']]
plt.figure(figsize=(16, 9))
sns.barplot(
data=df_multinode_offdef,
x="benchmark_type",
y="t_avg_usec",
dodge=True,
hue=df_multinode_offdef["msg_size_bytes"].astype(str),
)
# plt.yscale("log")
plt.title("Average Time (usec) per Benchmark Type and Message Size")
plt.ylabel("Average Time (usec)")
plt.xlabel("Benchmark Type")
plt.xticks(rotation=45)
plt.legend(title="Message Size (bytes)")
plt.tight_layout()
# plt.show()
plt.savefig("./plots/benchmark_avg_time_barplot.png", dpi=300)
plt.close()

64
results-and-plotting/python/scripts/plot_fast_group.py Normal file
View File

@@ -0,0 +1,64 @@
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np
from mpl_toolkits.mplot3d import Axes3D
data_file = "data/data-multi-defand100cflag.csv"
df_multinode = pd.read_csv(data_file, delimiter=',')
df_multinode_offdef = df_multinode[df_multinode['off_cache_flag'] == 100]
df_multinode_offdef = df_multinode_offdef[['benchmark_type','msg_size_bytes','t_avg_usec','proc_num']]
benchmarks = df_multinode_offdef['benchmark_type'].unique().tolist()
benchmarks = [x for x in benchmarks if x[-1] != 'v']
df_multinode_offdef = df_multinode_offdef[df_multinode_offdef['benchmark_type'].isin(
benchmarks)][df_multinode_offdef['msg_size_bytes'] > 1000]
fast_benchmarks = ["Allreduce","Bcast","Reduce","Reduce_scatter"]
df_multinode_offdef = df_multinode_offdef[df_multinode_offdef["benchmark_type"].isin(fast_benchmarks)]
plt.figure(figsize=(16, 9))
sns.barplot(
data=df_multinode_offdef,
x="benchmark_type",
y="t_avg_usec",
dodge=True,
hue=df_multinode_offdef["msg_size_bytes"].astype(str),
)
plt.ylim(0)
plt.title("Average Time (usec) per Benchmark Type and Message Size")
plt.ylabel("Average Time (usec)")
plt.xlabel("Benchmark Type")
plt.xticks(rotation=45)
plt.legend(title="Message Size (bytes)")
plt.tight_layout()
plt.savefig("./plots/fbenchmarks_avg_time_barplot.png", dpi=300)
plt.close()
df_allreduce= df_multinode_offdef[df_multinode_offdef["benchmark_type"]=="Allreduce"]
df_allreduce = df_allreduce[['msg_size_bytes','t_avg_usec','proc_num']]
df_allreduce = df_allreduce[df_allreduce['msg_size_bytes']>2**17]
pivot = df_allreduce.pivot(index="msg_size_bytes", columns="proc_num", values="t_avg_usec")
X = pivot.columns.values # proc_num
Y = pivot.index.values # msg_size_bytes
X, Y = np.meshgrid(X, Y)
Z = pivot.values
fig = plt.figure(figsize=(16, 9))
ax = fig.add_subplot(111, projection='3d')
surf = ax.plot_surface(X, Y, Z, cmap="viridis", edgecolor='k')
cbar = fig.colorbar(surf, ax=ax, shrink=0.6, pad=0.01, location='left')
cbar.set_label("Average Time (μs)")
ax.set_xlabel("Process Count")
ax.set_ylabel("Message Size (B)")
ax.set_zlabel("Average Time (μs)")
ax.set_title("Allreduce")
ax.set_xticks(pivot.columns.values) # use the actual process count values
ax.set_xticklabels(pivot.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.tight_layout()
plt.savefig("./plots/allreduce_surface.png", dpi=300)
plt.close()

66
results-and-plotting/python/scripts/plot_mid_group.py Normal file
View File

@@ -0,0 +1,66 @@
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np
from mpl_toolkits.mplot3d import Axes3D
data_file = "data/data-multi-defand100cflag.csv"
df_multinode = pd.read_csv(data_file, delimiter=',')
df_multinode_offdef = df_multinode[df_multinode['off_cache_flag'] == 100]
df_multinode_offdef = df_multinode_offdef[['benchmark_type','msg_size_bytes','t_avg_usec','proc_num']]
benchmarks = df_multinode_offdef['benchmark_type'].unique().tolist()
benchmarks = [x for x in benchmarks if x[-1] != 'v']
df_multinode_offdef = df_multinode_offdef[df_multinode_offdef['benchmark_type'].isin(
benchmarks)][df_multinode_offdef['msg_size_bytes'] > 1000]
# fast_benchmarks = ["Allreduce","Bcast","Reduce","Reduce_scatter"]
medium_benchmarks = ["Gather","Scatter"]
df_multinode_offdef = df_multinode_offdef[df_multinode_offdef["benchmark_type"].isin(medium_benchmarks)]
plt.figure(figsize=(16, 9))
sns.barplot(
data=df_multinode_offdef,
x="benchmark_type",
y="t_avg_usec",
dodge=True,
hue=df_multinode_offdef["msg_size_bytes"].astype(str),
)
plt.ylim(0)
plt.title("Average Time (usec) per Benchmark Type and Message Size")
plt.ylabel("Average Time (usec)")
plt.xlabel("Benchmark Type")
plt.xticks(rotation=45)
plt.legend(title="Message Size (bytes)")
plt.tight_layout()
plt.savefig("./plots/mbenchmarks_avg_time_barplot.png", dpi=300)
plt.close()
df_gather = df_multinode_offdef[df_multinode_offdef['benchmark_type']=='Gather']
df_gather = df_gather[['msg_size_bytes','t_avg_usec','proc_num']]
df_gather = df_gather[df_gather['msg_size_bytes']>2**17]
pivot = df_gather.pivot(index="msg_size_bytes", columns="proc_num", values="t_avg_usec")
X = pivot.columns.values # proc_num
Y = pivot.index.values # msg_size_bytes
X, Y = np.meshgrid(X, Y)
Z = pivot.values
fig = plt.figure(figsize=(16, 9))
ax = fig.add_subplot(111, projection='3d')
surf = ax.plot_surface(X, Y, Z, cmap="viridis", edgecolor='k')
cbar = fig.colorbar(surf, ax=ax, shrink=0.6, pad=0.01, location='left')
cbar.set_label("Average Time (μs)")
ax.set_xlabel("Process Count")
ax.set_ylabel("Message Size (B)")
ax.set_zlabel("Average Time (μs)")
ax.set_title("Gather")
ax.set_xticks(pivot.columns.values) # use the actual process count values
ax.set_xticklabels(pivot.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.tight_layout()
plt.savefig("./plots/gather_surface.png", dpi=300)
plt.close()

93
results-and-plotting/python/scripts/plot_slow_group.py Normal file
View File

@@ -0,0 +1,93 @@
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np
from mpl_toolkits.mplot3d import Axes3D
data_file = "data/data-multi-defand100cflag.csv"
df_multinode = pd.read_csv(data_file, delimiter=',')
df_multinode_offdef = df_multinode[df_multinode['off_cache_flag'] == 100]
df_multinode_offdef = df_multinode_offdef[['benchmark_type','msg_size_bytes','t_avg_usec','proc_num']]
benchmarks = df_multinode_offdef['benchmark_type'].unique().tolist()
benchmarks = [x for x in benchmarks if x[-1] != 'v']
df_multinode_offdef = df_multinode_offdef[df_multinode_offdef['benchmark_type'].isin(
benchmarks)][df_multinode_offdef['msg_size_bytes'] > 1000]
slow_benchmarks = ["Alltoall","Allgather"]
df_multinode_offdef = df_multinode_offdef[df_multinode_offdef["benchmark_type"].isin(slow_benchmarks)]
plt.figure(figsize=(16, 9))
sns.barplot(
data=df_multinode_offdef,
x="benchmark_type",
y="t_avg_usec",
dodge=True,
hue=df_multinode_offdef["msg_size_bytes"].astype(str),
)
plt.ylim(0)
plt.title("Average Time (usec) per Benchmark Type and Message Size")
plt.ylabel("Average Time (usec)")
plt.xlabel("Benchmark Type")
plt.xticks(rotation=45)
plt.legend(title="Message Size (bytes)")
plt.tight_layout()
plt.savefig("./plots/sbenchmarks_avg_time_barplot.png", dpi=300)
plt.close()
df_alltoall = df_multinode_offdef[df_multinode_offdef['benchmark_type']=='Alltoall']
df_alltoall = df_alltoall[['msg_size_bytes','t_avg_usec','proc_num']]
df_alltoall = df_alltoall[df_alltoall['msg_size_bytes']>2**17]
pivot = df_alltoall.pivot(index="msg_size_bytes", columns="proc_num", values="t_avg_usec")
X = pivot.columns.values # proc_num
Y = pivot.index.values # msg_size_bytes
X, Y = np.meshgrid(X, Y)
Z = pivot.values
fig = plt.figure(figsize=(16, 9))
ax = fig.add_subplot(111, projection='3d')
surf = ax.plot_surface(X, Y, Z, cmap="viridis", edgecolor='k')
cbar = fig.colorbar(surf, ax=ax, shrink=0.6, pad=0.01, location='left')
cbar.set_label("Average Time (μs)")
ax.set_xlabel("Process Count")
ax.set_ylabel("Message Size (B)")
ax.set_zlabel("Average Time (μs)")
ax.set_title("Alltoall")
ax.set_xticks(pivot.columns.values) # use the actual process count values
ax.set_xticklabels(pivot.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.tight_layout()
plt.savefig("./plots/alltoall_surface.png", dpi=300)
plt.close()
df_allgather = df_multinode_offdef[df_multinode_offdef['benchmark_type']=='Allgather']
df_allgather = df_allgather[['msg_size_bytes','t_avg_usec','proc_num']]
df_allgather = df_allgather[df_allgather['msg_size_bytes']>2**17]
pivot = df_allgather.pivot(index="msg_size_bytes", columns="proc_num", values="t_avg_usec")
X = pivot.columns.values # proc_num
Y = pivot.index.values # msg_size_bytes
X, Y = np.meshgrid(X, Y)
Z = pivot.values
fig = plt.figure(figsize=(16, 9))
ax = fig.add_subplot(111, projection='3d')
surf = ax.plot_surface(X, Y, Z, cmap="viridis", edgecolor='k')
cbar = fig.colorbar(surf, ax=ax, shrink=0.6, pad=0.01, location='left')
cbar.set_label("Average Time (μs)")
ax.set_xlabel("Process Count")
ax.set_ylabel("Message Size (B)")
ax.set_zlabel("Average Time (μs)")
ax.set_title("Allgather")
ax.set_xticks(pivot.columns.values) # use the actual process count values
ax.set_xticklabels(pivot.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.tight_layout()
plt.savefig("./plots/allgather_surface.png", dpi=300)
plt.close()