1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
|
import csv
import pickle
import matplotlib.pyplot as plt
import multiprocessing
import numpy as np
import os
import re
import subprocess
from benchmark import Benchmark
from common_targets import common_targets
cmd = ("perf stat -d -x\; -e cpu-clock,cache-references,cache-misses,cycles,"
"instructions,branches,faults,migrations "
"build/cache-{}{} {} 100 8 1000000")
class Benchmark_Falsesharing( Benchmark ):
def __init__(self):
self.name = "falsesharing"
self.descrition = """This benchmarks makes small allocations and writes
to them multiple times. If the allocated objects are
on the same cache line the writes will be expensive because
of cache thrashing.""",
self.targets = common_targets
self.nthreads = range(1, multiprocessing.cpu_count() * 2 + 1)
self.results = {"args" : {"nthreads" : self.nthreads},
"targets" : self.targets,
"thrash": {},
"scratch": {}}
def prepare(self, verbose=False):
req = ["build/cache-thrash", "build/cache-scratch"]
for r in req:
if not os.path.isfile(r):
print(r, "not found")
return False
if not os.access(r, os.X_OK):
print(r, "not executable")
return False
if verbose:
print(r, "found and executable.")
return True
def run(self, verbose=False, runs=3):
for run in range(1, runs + 1):
print(str(run) + ". run")
n = len(self.nthreads)
for i, threads in enumerate(list(range(1, n + 1)) * 2):
print(i + 1, "of", n*2, "\r", end='')
# run cmd for each target
for tname, t in self.targets.items():
result = {}
os.environ["LD_PRELOAD"] = t[1]
for bench in ["thrash", "scratch"]:
target_cmd = cmd.format(bench, t[0], threads).split(" ")
if verbose:
print("\n" + tname, t, "\n", " ".join(target_cmd), "\n")
p = subprocess.run(target_cmd,
env=os.environ,
stderr=subprocess.PIPE,
stdout=subprocess.PIPE,
universal_newlines=True)
output = str(p.stdout)
err = str(p.stderr)
if p.returncode != 0:
print("\n" + " ".join(target_cmd), "exited with",
p.returncode, ".\n Aborting Benchmark.")
print(tname, t)
print(output)
print(p.stdout)
return False
if "ERROR: ld.so" in output:
print("\nPreloading of", t[1], "failed for", tname,
".\n Aborting Benchmark.")
print(output)
return False
time = float(re.search("(\d*\.\d*)", output)[1])
result["time"] = time
# Handle perf output
csvreader = csv.reader(err.splitlines()[1:], delimiter=';')
for row in csvreader:
result[row[2].replace("\\", "")] = row[0].replace("\\", "")
key = (tname, threads)
if not key in self.results[bench]:
self.results[bench][key] = [result]
else:
self.results[bench][key].append(result)
print()
return True
def summary(self):
# Speedup thrash
nthreads = self.results["args"]["nthreads"]
targets = self.results["targets"]
y_mapping = {v : i for i, v in enumerate(nthreads)}
for bench in ["thrash", "scratch"]:
for target in targets:
y_vals = [0] * len(nthreads)
single_threaded = np.mean([m["time"] for m in self.results[bench][(target, 1)]])
for mid, measures in self.results[bench].items():
if mid[0] == target:
l1_load_misses = []
d = []
for m in measures:
d.append(m["time"])
misses = float(m["L1-dcache-load-misses"])
loads = float(m["L1-dcache-loads"])
l1_load_misses.append(misses/loads)
y_vals[y_mapping[mid[1]]] = single_threaded / np.mean(d)
s = "{} {} {}: {:.3f}%".format(bench, target, mid[1], np.mean(l1_load_misses)*100)
print(s)
plt.plot(nthreads, y_vals, marker='.', linestyle='-', label=target)
plt.legend()
plt.xlabel("threads")
plt.ylabel("speedup")
plt.title(bench)
plt.savefig(self.name + "." + bench + ".png")
plt.clf()
falsesharing= Benchmark_Falsesharing()
|
