Mercurial > cgi-bin > hgwebdir.cgi > VMS > 2__runs_and_data

comparison scripts/overhead.py @ 1:75c55af8338f

find changesets by author, revision, files, or words in the commit message
Overhead.py: Script to generate the results of the exec_vs_task_size benchmarks
author Merten Sach <msach@mailbox.tu-berlin.de>
date Fri, 09 Dec 2011 15:09:08 +0100
parents
children 8323aae8c303
comparison
equal deleted inserted replaced
-1:000000000000 0:30682dd71414
1 #! /usr/bin/env python
2 # -*- coding: utf-8 -*-
3
4 import sys
5 from re import match, search
6 from datetime import datetime
7 from subprocess import call,Popen,PIPE
8
9 """
10 This script generates a graph that represents the overhead
11
12 involved in synchronisation operations
13 """
14
15 usage="""
16 This runs the exec time vs task size in three levels of loop nest. The outer most iterates through
17 a selection of numbers-of-thread. For each of those, the next lever iterates over a number of work-loops-per-task
18 values. The innermost repeats several times and chooses the best.
19 Finally, it generates an output file for each value of number-of-threads that a companion gluplot script turns
20 into a .eps graph.
21 It is expected that the output directory's path is meaningful, such as machine-name, date, and so on
22 Usage:
23 overhead.py [executable binary] [path to output dir]
24 """
25
26 NUM_CORES = 4 #Number of Cores the code was compiled for
27 ITERS_PER_TASK_TABLE = [2, 5, 10, 20, 40, 80, 160, 320, 640] #Number of iterations of inner loop
28 TASKS_PER_THREAD = 30000 #Number of interations of outer loop
29 TOTAL_THREADS_TABLE = [8, 32, 128, 512]
30
31 def getNumber(line):
32 match_obj = search("(\d+\.?\d*)", line)
33 if match_obj != None:
34 return match_obj.groups()[0]
35 else:
36 raise ValueError
37
38 if len(sys.argv) != 3:
39 print usage
40 sys.exit(0)
41
42 cmd=sys.argv[1]
43 try:
44 f = open(cmd)
45 except IOError:
46 print "Please provide a valid executable."
47 f.close()
48 sys.exit(1)
49 finally:
50 f.close()
51
52 output_dir_path = sys.argv[2]
53
54 #===================================================================
55 # Done with parsing cmd line inputs, start doing the runs
56 #
57
58 for totalThreads in TOTAL_THREADS_TABLE:
59 print "\nDoing run with %d threads" % totalThreads
60 output = "%s/%d_thds__o%d__perfCtrs.meas" % (output_dir_path, totalThreads, TASKS_PER_THREAD)
61 print "output file: %s" % output
62 threadsPerCore = totalThreads/NUM_CORES
63 array_of_results = []
64 for workload_iterations_in_task in ITERS_PER_TASK_TABLE:
65 print "Run for %s workload iterations in a task" % workload_iterations_in_task
66 results = []
67 for run in range(5):
68 print "Run %d" % run,
69 program_output = Popen("%s -t %d -i %d -o %d" % (cmd,
70 totalThreads,
71 workload_iterations_in_task,
72 TASKS_PER_THREAD),
73 stdout=PIPE, stderr=None, shell=True).stdout.read()
74 #parse arguments for
75 for line in program_output.split("\n"):
76 if match("^Sum across threads of work cycles:", line) != None:
77 total_workcycles = int(getNumber(line))
78 if match("^Total Execution Cycles:", line) != None:
79 total_exe_cycles = int(getNumber(line))
80 if match("^ExeCycles/WorkCycles Ratio", line) != None:
81 exeCycles_workCycles_ratio = float(getNumber(line))
82 results.append({"total_workcycles" : total_workcycles,
83 "total_exe_cycles" : total_exe_cycles,
84 "exeCycles_workCycles_ratio" : exeCycles_workCycles_ratio})
85 print "ratio %f" % exeCycles_workCycles_ratio
86 array_of_results.append(results)
87
88
89 #open gnuplot output
90 try:
91 gnuplot_output = open(output,"w")
92 except IOError:
93 print "Cannot open output file %s" % output
94 sys.exit(1)
95
96 table_header = "# %20s\t%20s\t%20s\t%20s\t%20s\t%20s\t%20s\t%20s\n" % (
97 "<iters per task>",
98 "<total exe cycles>",
99 "<total work cyc>",
100 "<one task cyc>",
101 "<total overhead cyc>",
102 "<num syncs>",
103 "<overhead per Sync cyc>",
104 "<Exe/Work ratio>")
105
106 #write header to file
107 gnuplot_output.writelines(["# Output file name: %s\n" % output,
108 "# Date of Run: %s\n" % str(datetime.now()),
109 "# Number of Cores: %d\n" % NUM_CORES,
110 "# Number of Threads: %f per Core, %d total\n" % (threadsPerCore, totalThreads),
111 table_header,
112 "# " + (len(table_header)-3)*"-" + "\n"])
113
114 #Now print the results out
115 idx = -1
116 for workload_iterations_in_task in ITERS_PER_TASK_TABLE:
117 idx += 1
118 results = array_of_results[idx]
119
120 #take shortest run
121 results.sort(lambda x,y: cmp(x["total_exe_cycles"],y["total_exe_cycles"]))
122 total_workcycles = results[0]["total_workcycles"]
123 total_exe_cycles = results[0]["total_exe_cycles"]
124 exeCycles_workCycles_ratio = results[0]["exeCycles_workCycles_ratio"]
125
126 #Calculate numbers
127 overhead = total_exe_cycles - total_workcycles
128 total_syncs = totalThreads * TASKS_PER_THREAD * 2
129 overhead_per_sync = float(overhead) / float(total_syncs)
130 cycles_of_task = float(total_workcycles) / float(TASKS_PER_THREAD * totalThreads)
131 overhead_per_core = float(overhead) / NUM_CORES
132 workcycles_per_core = total_workcycles / NUM_CORES
133
134 gnuplot_output.write("%20d\t%20d\t%20d\t%20f\t%20d\t%20d\t%20f\t%20f\n" % (
135 workload_iterations_in_task,
136 total_exe_cycles,
137 total_workcycles,
138 cycles_of_task,
139 overhead,
140 total_syncs,
141 overhead_per_sync,
142 exeCycles_workCycles_ratio))
143
144 gnuplot_output.close();