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

annotate scripts/overhead.py @ 3:8323aae8c303

find changesets by author, revision, files, or words in the commit message
plot script rewritten in python and changed labeling
author Merten Sach <msach@mailbox.tu-berlin.de>
date Mon, 12 Dec 2011 20:23:22 +0100
parents 75c55af8338f
children
rev   line source
msach@1 1 #! /usr/bin/env python
msach@1 2 # -*- coding: utf-8 -*-
msach@1 3
msach@1 4 import sys
msach@1 5 from re import match, search
msach@1 6 from datetime import datetime
msach@1 7 from subprocess import call,Popen,PIPE
msach@1 8
msach@1 9 """
msach@1 10 This script generates a graph that represents the overhead
msach@1 11
msach@1 12 involved in synchronisation operations
msach@1 13 """
msach@1 14
msach@1 15 usage="""
msach@1 16 This runs the exec time vs task size in three levels of loop nest. The outer most iterates through
msach@1 17 a selection of numbers-of-thread. For each of those, the next lever iterates over a number of work-loops-per-task
msach@1 18 values. The innermost repeats several times and chooses the best.
msach@1 19 Finally, it generates an output file for each value of number-of-threads that a companion gluplot script turns
msach@1 20 into a .eps graph.
msach@1 21 It is expected that the output directory's path is meaningful, such as machine-name, date, and so on
msach@1 22 Usage:
msach@1 23 overhead.py [executable binary] [path to output dir]
msach@1 24 """
msach@1 25
msach@1 26 NUM_CORES = 4 #Number of Cores the code was compiled for
msach@1 27 ITERS_PER_TASK_TABLE = [2, 5, 10, 20, 40, 80, 160, 320, 640] #Number of iterations of inner loop
msach@1 28 TASKS_PER_THREAD = 30000 #Number of interations of outer loop
msach@1 29 TOTAL_THREADS_TABLE = [8, 32, 128, 512]
msach@1 30
msach@1 31 def getNumber(line):
msach@1 32 match_obj = search("(\d+\.?\d*)", line)
msach@1 33 if match_obj != None:
msach@1 34 return match_obj.groups()[0]
msach@1 35 else:
msach@1 36 raise ValueError
msach@1 37
msach@1 38 if len(sys.argv) != 3:
msach@1 39 print usage
msach@1 40 sys.exit(0)
msach@1 41
msach@1 42 cmd=sys.argv[1]
msach@1 43 try:
msach@1 44 f = open(cmd)
msach@1 45 except IOError:
msach@1 46 print "Please provide a valid executable."
msach@1 47 f.close()
msach@1 48 sys.exit(1)
msach@1 49 finally:
msach@1 50 f.close()
msach@1 51
msach@1 52 output_dir_path = sys.argv[2]
msach@1 53
msach@1 54 #===================================================================
msach@1 55 # Done with parsing cmd line inputs, start doing the runs
msach@1 56 #
msach@1 57
msach@1 58 for totalThreads in TOTAL_THREADS_TABLE:
msach@1 59 print "\nDoing run with %d threads" % totalThreads
msach@1 60 output = "%s/%d_thds__o%d__perfCtrs.meas" % (output_dir_path, totalThreads, TASKS_PER_THREAD)
msach@1 61 print "output file: %s" % output
msach@1 62 threadsPerCore = totalThreads/NUM_CORES
msach@1 63 array_of_results = []
msach@1 64 for workload_iterations_in_task in ITERS_PER_TASK_TABLE:
msach@1 65 print "Run for %s workload iterations in a task" % workload_iterations_in_task
msach@1 66 results = []
msach@1 67 for run in range(5):
msach@1 68 print "Run %d" % run,
msach@1 69 program_output = Popen("%s -t %d -i %d -o %d" % (cmd,
msach@1 70 totalThreads,
msach@1 71 workload_iterations_in_task,
msach@1 72 TASKS_PER_THREAD),
msach@1 73 stdout=PIPE, stderr=None, shell=True).stdout.read()
msach@1 74 #parse arguments for
msach@1 75 for line in program_output.split("\n"):
msach@1 76 if match("^Sum across threads of work cycles:", line) != None:
msach@1 77 total_workcycles = int(getNumber(line))
msach@1 78 if match("^Total Execution Cycles:", line) != None:
msach@1 79 total_exe_cycles = int(getNumber(line))
msach@1 80 if match("^ExeCycles/WorkCycles Ratio", line) != None:
msach@1 81 exeCycles_workCycles_ratio = float(getNumber(line))
msach@1 82 results.append({"total_workcycles" : total_workcycles,
msach@1 83 "total_exe_cycles" : total_exe_cycles,
msach@1 84 "exeCycles_workCycles_ratio" : exeCycles_workCycles_ratio})
msach@1 85 print "ratio %f" % exeCycles_workCycles_ratio
msach@1 86 array_of_results.append(results)
msach@1 87
msach@1 88
msach@1 89 #open gnuplot output
msach@1 90 try:
msach@1 91 gnuplot_output = open(output,"w")
msach@1 92 except IOError:
msach@1 93 print "Cannot open output file %s" % output
msach@1 94 sys.exit(1)
msach@1 95
msach@1 96 table_header = "# %20s\t%20s\t%20s\t%20s\t%20s\t%20s\t%20s\t%20s\n" % (
msach@1 97 "<iters per task>",
msach@1 98 "<total exe cycles>",
msach@1 99 "<total work cyc>",
msach@1 100 "<one task cyc>",
msach@1 101 "<total overhead cyc>",
msach@1 102 "<num syncs>",
msach@1 103 "<overhead per Sync cyc>",
msach@1 104 "<Exe/Work ratio>")
msach@1 105
msach@1 106 #write header to file
msach@1 107 gnuplot_output.writelines(["# Output file name: %s\n" % output,
msach@1 108 "# Date of Run: %s\n" % str(datetime.now()),
msach@1 109 "# Number of Cores: %d\n" % NUM_CORES,
msach@1 110 "# Number of Threads: %f per Core, %d total\n" % (threadsPerCore, totalThreads),
msach@1 111 table_header,
msach@1 112 "# " + (len(table_header)-3)*"-" + "\n"])
msach@1 113
msach@1 114 #Now print the results out
msach@1 115 idx = -1
msach@1 116 for workload_iterations_in_task in ITERS_PER_TASK_TABLE:
msach@1 117 idx += 1
msach@1 118 results = array_of_results[idx]
msach@1 119
msach@1 120 #take shortest run
msach@1 121 results.sort(lambda x,y: cmp(x["total_exe_cycles"],y["total_exe_cycles"]))
msach@1 122 total_workcycles = results[0]["total_workcycles"]
msach@1 123 total_exe_cycles = results[0]["total_exe_cycles"]
msach@3 124 #exeCycles_workCycles_ratio = results[0]["exeCycles_workCycles_ratio"]
msach@1 125
msach@1 126 #Calculate numbers
msach@1 127 overhead = total_exe_cycles - total_workcycles
msach@1 128 total_syncs = totalThreads * TASKS_PER_THREAD * 2
msach@1 129 overhead_per_sync = float(overhead) / float(total_syncs)
msach@1 130 cycles_of_task = float(total_workcycles) / float(TASKS_PER_THREAD * totalThreads)
msach@1 131 overhead_per_core = float(overhead) / NUM_CORES
msach@1 132 workcycles_per_core = total_workcycles / NUM_CORES
msach@3 133
msach@3 134 exeCycles_workCycles_ratio = float(total_workcycles+float(overhead)/2)/float(total_workcycles)
msach@1 135
msach@1 136 gnuplot_output.write("%20d\t%20d\t%20d\t%20f\t%20d\t%20d\t%20f\t%20f\n" % (
msach@1 137 workload_iterations_in_task,
msach@1 138 total_exe_cycles,
msach@1 139 total_workcycles,
msach@1 140 cycles_of_task,
msach@1 141 overhead,
msach@1 142 total_syncs,
msach@1 143 overhead_per_sync,
msach@1 144 exeCycles_workCycles_ratio))
msach@1 145
msach@1 146 gnuplot_output.close();