303 lines
13 KiB
Python
303 lines
13 KiB
Python
#! /bin/python3
|
|
|
|
import os
|
|
import re
|
|
from dataclasses import dataclass
|
|
import json
|
|
import matplotlib.pyplot as plt
|
|
import numpy
|
|
from scipy.io import mmread
|
|
import math
|
|
import itertools
|
|
import argparse
|
|
|
|
perf_stats = ["task-clock:u", "page-faults:u", "cycles:u", "instructions:u",
|
|
"branches:u", "branch-misses:u", "L1-dcache-loads:u",
|
|
"L1-dcache-load-misses:u", "LLC-loads:u", "LLC-load-misses:u",
|
|
"L1-icache-loads:u", "L1-icache-load-misses:u", "dTLB-loads:u",
|
|
"dTLB-load-misses:u", "iTLB-loads:u", "iTLB-load-misses:u"]
|
|
|
|
def read_stats(filename):
|
|
# Split filename into each input parameter.
|
|
filename_split = os.path.splitext(filename)[0].split('_')
|
|
s = {
|
|
"solver": filename_split[0],
|
|
"linalg": filename_split[1],
|
|
"file": filename_split[2],
|
|
"time_steps": int(filename_split[3]),
|
|
"cores": int(re.match(r'\d+', filename_split[4]).group(0)),
|
|
"cpu": filename_split[5] }
|
|
|
|
# Take filename and read SchurComplement matrix from corresponding dir.
|
|
print(os.path.splitext(filename)[0])
|
|
m = mmread(os.path.splitext(filename)[0] + "/SchurComplement.mm")
|
|
s["cols"] = m.get_shape()[0]
|
|
s["nonzero"] = m.count_nonzero()
|
|
s["entries"] = m.get_shape()[0] * m.get_shape()[1]
|
|
s["trace"] = m.trace()
|
|
|
|
# Compute ratio.
|
|
s["% nonzero entry"] = s["nonzero"] / s["entries"]
|
|
s["trace over cols"] = s["trace"] / s["cols"]
|
|
|
|
# Obtain stats from perf in each log.
|
|
with open(filename) as file:
|
|
for line in file:
|
|
for perf_stat in perf_stats:
|
|
regex = r'^\W*([\d+(,|\.)?]+)\W*.*' + perf_stat
|
|
data = re.search(regex, line)
|
|
if data is None:
|
|
continue
|
|
match perf_stat:
|
|
case "task-clock:u":
|
|
s[perf_stat[:-2] + " (msec)"] = float(data.group(1).replace(',', ''))
|
|
case "page-faults:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "cycles:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "instructions:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "branches:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "branch-misses:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "L1-dcache-loads:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "L1-dcache-load-misses:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "LLC-loads:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "LLC-load-misses:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "L1-icache-loads:u":
|
|
print("l1_icache_loads not expected! Exiting...")
|
|
exit(1)
|
|
case "L1-icache-load-misses:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "dTLB-loads:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "dTLB-load-misses:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "iTLB-loads:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case "iTLB-load-misses:u":
|
|
s[perf_stat[:-2]] = int(data.group(1).replace(',', ''))
|
|
case _:
|
|
print("Unexpected case! Exiting...")
|
|
exit(1)
|
|
|
|
# Derive stats.
|
|
if s["cycles"] != 0:
|
|
s["insn per cycle"] = s["instructions"] / s["cycles"]
|
|
else:
|
|
s["insn per cycle"] = None
|
|
s["branch-miss-rate"] = s["branch-misses"] / s["branches"]
|
|
s["L1-dcache-miss-rate"] = s["L1-dcache-load-misses"] / s["L1-dcache-loads"]
|
|
s["LLC-miss-rate"] = s["LLC-load-misses"] / s["LLC-loads"]
|
|
s["dTLB-miss-rate"] = s["dTLB-load-misses"] / s["dTLB-loads"]
|
|
s["iTLB-miss-rate"] = s["iTLB-load-misses"] / s["iTLB-loads"]
|
|
|
|
print(s)
|
|
return s
|
|
|
|
def box_plot(ax, stat_list, x_name, y_name):
|
|
data = accumulate_to_dict(stat_list, x_name, y_name)
|
|
|
|
for value in data.values():
|
|
if type(value[0]) is str:
|
|
print(f"{y_name} values are strings... skipping box plot...")
|
|
return
|
|
|
|
print("Plotted data: " + str(data))
|
|
|
|
ax.boxplot([numpy.array(y_axis) for y_axis in data.values()])
|
|
ax.set_xticklabels(data.keys())
|
|
ax.set_ylabel(y_name)
|
|
|
|
def scatter_plot(ax, stat_list, x_name, y_name, color_name, size_name = None):
|
|
x_data = accumulate_to_dict(stat_list, color_name, x_name)
|
|
y_data = accumulate_to_dict(stat_list, color_name, y_name)
|
|
if size_name is not None:
|
|
size_data = accumulate_to_dict(stat_list, color_name, size_name)
|
|
|
|
print(x_data)
|
|
print(y_data)
|
|
if size_name is not None:
|
|
print(size_data)
|
|
|
|
for group in x_data.keys():
|
|
print(f"Plotted x data for {group}: " + str(x_data[group]))
|
|
print(f"Plotted y data for {group}: " + str(y_data[group]))
|
|
ax.scatter(x_data[group], y_data[group], label = group, alpha = 0.3)
|
|
|
|
#ax.legend()
|
|
#ax.set_xlabel(x_name)
|
|
ax.set_ylabel(y_name)
|
|
ax.grid(True)
|
|
|
|
def line_plot(ax, stat_list, x_name, y_name, color_name = None, marker_name = None):
|
|
x_data = accumulate_to_dict(stat_list, color_name, x_name, marker_name)
|
|
y_data = accumulate_to_dict(stat_list, color_name, y_name, marker_name)
|
|
|
|
for group in x_data.keys():
|
|
sorted_data = [(x, y) for (x, y) in zip(x_data[group], y_data[group])]
|
|
sorted_data.sort()
|
|
x_data[group] = [x for (x, y) in sorted_data]
|
|
y_data[group] = [y for (x, y) in sorted_data]
|
|
print(f"Plotted x data for {group}: " + str(x_data[group]))
|
|
print(f"Plotted y data for {group}: " + str(y_data[group]))
|
|
ax.plot(x_data[group], y_data[group], label = group)
|
|
|
|
ax.set_ylabel(y_name)
|
|
ax.grid(True)
|
|
|
|
def basic_plot(ax, stat_list, x_name, y_name):
|
|
xs = accumulate_to_list(stat_list, x_name)
|
|
ys = accumulate_to_list(stat_list, y_name)
|
|
print("Plotted x data: " + str(xs))
|
|
print("Plotted y data: " + str(ys))
|
|
|
|
ax.plot(xs, ys, 'o')
|
|
|
|
ax.legend()
|
|
ax.set_xlabel(x_name)
|
|
ax.set_ylabel(y_name)
|
|
ax.grid(True)
|
|
|
|
def accumulate_to_dict(stat_list, key_name, value_name, key2_name = None):
|
|
key_list = [stat[key_name] for stat in stat_list]
|
|
value_list = [stat[value_name] for stat in stat_list]
|
|
|
|
if key2_name is not None:
|
|
key2_list = [stat[key2_name] for stat in stat_list]
|
|
data = {pair: list() for pair in
|
|
list(itertools.product(list(dict.fromkeys(key_list).keys()),
|
|
list(dict.fromkeys(key2_list).keys())))}
|
|
for key, key2, value in zip(key_list, key2_list, value_list):
|
|
data[(key, key2)].append(value)
|
|
else:
|
|
data = {key: list() for key in dict.fromkeys(key_list)}
|
|
for key, value in zip(key_list, value_list):
|
|
data[key].append(value)
|
|
|
|
return data
|
|
|
|
def accumulate_to_list(stat_list, name):
|
|
lst = [stat[name] for stat in stat_list]
|
|
return lst
|
|
|
|
def show_visualizations(rows, plot, size_multiplier, stat_list, x_name, y_name = None,
|
|
color_name = None, marker_name = None, filter_list = []):
|
|
# Filter stats (removing values such as a specific CPU type)
|
|
stat_list = [stat for stat in stat_list
|
|
if len([stat[key] for key in stat.keys()
|
|
if stat[key] in filter_list])
|
|
== 0]
|
|
|
|
if y_name is None or y_name == "all":
|
|
# Filter y names (removing measurements such as cycles)
|
|
y_names = [y_name for y_name in stat_list[0].keys() if y_name != x_name
|
|
and y_name != color_name
|
|
and y_name != marker_name
|
|
and y_name not in filter_list]
|
|
|
|
fig, axes = plt.subplots(rows, int(math.ceil(len(y_names) / rows)),
|
|
figsize = (16 * size_multiplier, 9 * size_multiplier))
|
|
match plot:
|
|
case "box":
|
|
for i, y in enumerate(y_names):
|
|
box_plot(axes[i % rows][int(i / rows)], stat_list, x_name, y)
|
|
case "line":
|
|
for i, y in enumerate(y_names):
|
|
line_plot(axes[i % rows][int(i / rows)], stat_list, x_name,
|
|
y, color_name, marker_name)
|
|
case _:
|
|
print("Invalid plot type! Exiting...")
|
|
exit(1)
|
|
|
|
handles, labels = axes[i % rows][int(i / rows)].get_legend_handles_labels()
|
|
else:
|
|
fig, ax = plt.subplots()
|
|
match plot:
|
|
case "box":
|
|
box_plot(ax, stat_list, x_name, y_name)
|
|
case "line":
|
|
line_plot(ax, stat_list, x_name, y_name, color_name, marker_name)
|
|
case _:
|
|
print("Invalid plot type! Exiting...")
|
|
exit(1)
|
|
|
|
handles, labels = ax.get_legend_handles_labels()
|
|
|
|
if y_name is None or y_name == "all":
|
|
y_name = "all"
|
|
fontsize = 'xx-large'
|
|
else:
|
|
fontsize = 'x-small'
|
|
|
|
title = f"{plot}_plot_of_{y_name}_vs_{x_name}_by_{color_name}_and_{marker_name}_excluding_{filter_list}"
|
|
|
|
fig.suptitle(title, fontsize = fontsize)
|
|
fig.legend(handles, labels, fontsize = fontsize)
|
|
fig.supxlabel(x_name, fontsize = fontsize)
|
|
|
|
plt.savefig(title + ".png", dpi = 100)
|
|
plt.show()
|
|
|
|
save_file = "stats.json"
|
|
samples_dir = "Samples/"
|
|
# save_file = "stats_avg_10.json"
|
|
# samples_dir = "Samples_Avg_10/"
|
|
# save_file = "tmp.json"
|
|
# samples_dir = "Old_Samples_Avg_10"
|
|
|
|
def main():
|
|
stat_list = list()
|
|
|
|
# Get stats and accumulate into stat_list
|
|
if os.path.isfile(save_file):
|
|
with open(save_file, 'r') as file:
|
|
stat_list = json.load(file)
|
|
# print(*stat_list, sep='\n')
|
|
else:
|
|
os.chdir('./' + samples_dir)
|
|
for filename in os.listdir(os.getcwd()):
|
|
if os.path.splitext(filename)[1] == ".log":
|
|
stat_list.append(read_stats(filename))
|
|
print(*stat_list, sep='\n')
|
|
os.chdir('../')
|
|
with open(save_file, 'w') as file:
|
|
json.dump(stat_list, file, indent = 2)
|
|
|
|
# print(*stat_list, sep='\n')
|
|
|
|
possible_args = [keys for keys in stat_list[0].keys()]
|
|
parser = argparse.ArgumentParser(epilog = 'Possible arguments are: ' + str(possible_args))
|
|
parser.add_argument('-r', '--rows', help = 'Number of rows to display. Will split graphs into columns to evenly fill rows. Does not do anything if -y is not "all".', default = 0)
|
|
parser.add_argument('-p', '--plot', help = 'The type of plot. Only "box" and "line" are supported.')
|
|
parser.add_argument('-x', help = 'Name of the x axis. See below for possible arguments.')
|
|
parser.add_argument('-y', help = 'Name of the y axis. Can be "all" or omitted entirely to display all possible y values. See below for other possible arguments.')
|
|
parser.add_argument('-g1', '--group1', help = 'Name of the first group to group by. Does not need to be specified.')
|
|
parser.add_argument('-g2', '--group2', help = 'Name of the second group to group by. Does not need to be specified (and does not matter for box plots).')
|
|
parser.add_argument('-f', '--filter', nargs = '+', help = 'Names and string values to filter out of the visualization. Can be multiple space-separate arguments. Solver filters are "MueLu", "CG", "GMRES", and "ML". CPU filters are "Xeon,4216" and "Xeon,E5-2683"4216".', default = [])
|
|
parser.add_argument('-s', '--size', help = 'Multiplier for the resultant figure size if the visualization is too difficult to view. Default is 4.', default = 4)
|
|
|
|
args = parser.parse_args()
|
|
|
|
print(args)
|
|
|
|
if args.plot == "line" and args.group1 is None:
|
|
print("Line plot requires at least one group! Exiting...")
|
|
exit(1)
|
|
|
|
# Plot
|
|
show_visualizations(int(args.rows), args.plot, int(args.size), stat_list, args.x, args.y, args.group1, args.group2, args.filter)
|
|
#show_visualizations(4, "line", stat_list, "cols", None, "solver", "cpu", ["Xeon,E5-2683", "time_steps", "cpu", "linalg", "file", "cores"])
|
|
#show_visualizations(4, "line", stat_list, "cols", None, "solver", "cpu", ["time_steps", "cpu", "linalg", "file", "cores"])
|
|
#show_visualizations(4, "line", stat_list, "LLC-miss-rate", "branch-miss-rate", "solver", "cpu")
|
|
#show_visualizations(4, "box", stat_list, "solver", "cycles", None, None, ["Xeon,E5-2683", 0, None])
|
|
#show_visualizations(4, "box", stat_list, "solver", None, None, None, [0, None])
|
|
|
|
if __name__ == '__main__':
|
|
main()
|