1#!/usr/bin/env python 2import os,sys 3sys.path.append(os.path.join(os.environ['PETSC_DIR'], 'config')) 4sys.path.append(os.getcwd()) 5from builder2 import buildExample 6from benchmarkBatch import generateBatchScript 7 8class PETSc(object): 9 def __init__(self): 10 return 11 12 def dir(self): 13 '''Return the root directory for the PETSc tree (usually $PETSC_DIR)''' 14 # This should search for a valid PETSc 15 return os.environ['PETSC_DIR'] 16 17 def arch(self): 18 '''Return the PETSc build label (usually $PETSC_ARCH)''' 19 # This should be configurable 20 return os.environ['PETSC_ARCH'] 21 22 def mpiexec(self): 23 '''Return the path for the mpi launch executable''' 24 mpiexec = os.path.join(self.dir(), self.arch(), 'bin', 'mpiexec') 25 if not os.path.isfile(mpiexec): 26 return None 27 return mpiexec 28 29 def example(self, num): 30 '''Return the path to the executable for a given example number''' 31 return os.path.join(self.dir(), self.arch(), 'lib', 'ex'+str(num)+'-obj', 'ex'+str(num)) 32 33 def source(self, library, num, filenametail): 34 '''Return the path to the sources for a given example number''' 35 d = os.path.join(self.dir(), 'src', library.lower(), 'examples', 'tutorials') 36 name = 'ex'+str(num) 37 sources = [] 38 for f in os.listdir(d): 39 if f == name+'.c': 40 sources.insert(0, f) 41 elif f.startswith(name) and f.endswith(filenametail): 42 sources.append(f) 43 return map(lambda f: os.path.join(d, f), sources) 44 45class PETScExample(object): 46 def __init__(self, library, num, **defaultOptions): 47 self.petsc = PETSc() 48 self.library = library 49 self.num = num 50 self.opts = defaultOptions 51 return 52 53 @staticmethod 54 def runShellCommand(command, cwd = None, log = True): 55 import subprocess 56 57 Popen = subprocess.Popen 58 PIPE = subprocess.PIPE 59 if log: print 'Executing: %s\n' % (command,) 60 pipe = Popen(command, cwd=cwd, stdin=None, stdout=PIPE, stderr=PIPE, bufsize=-1, shell=True, universal_newlines=True) 61 (out, err) = pipe.communicate() 62 ret = pipe.returncode 63 return (out, err, ret) 64 65 def optionsToString(self, **opts): 66 '''Convert a dictionary of options to a command line argument string''' 67 a = [] 68 for key,value in opts.iteritems(): 69 if value is None: 70 a.append('-'+key) 71 else: 72 a.append('-'+key+' '+str(value)) 73 return ' '.join(a) 74 75 def run(self, numProcs = 1, log = True, **opts): 76 cmd = '' 77 if self.petsc.mpiexec() is not None: 78 cmd += self.petsc.mpiexec() + ' ' 79 numProcs = os.environ.get('NUM_RANKS', numProcs) 80 cmd += ' -n ' + str(numProcs) + ' ' 81 if os.environ.has_key('PE_HOSTFILE'): 82 cmd += ' -hostfile hostfile ' 83 cmd += ' '.join([self.petsc.example(self.num), self.optionsToString(**self.opts), self.optionsToString(**opts)]) 84 if 'batch' in opts and opts['batch']: 85 del opts['batch'] 86 filename = generateBatchScript(self.num, numProcs, 120, ' '+self.optionsToString(**self.opts)+' '+self.optionsToString(**opts)) 87 # Submit job 88 out, err, ret = self.runShellCommand('qsub -q gpu '+filename, log = log) 89 if ret: 90 print err 91 print out 92 else: 93 out, err, ret = self.runShellCommand(cmd, log = log) 94 if ret: 95 print err 96 print out 97 return out 98 99def processSummary(moduleName, defaultStage, eventNames, times, events): 100 '''Process the Python log summary into plot data''' 101 m = __import__(moduleName) 102 reload(m) 103 # Total Time 104 times.append(m.Time[0]) 105 # Particular events 106 for name in eventNames: 107 if name.find(':') >= 0: 108 stageName, name = name.split(':', 1) 109 stage = getattr(m, stageName) 110 else: 111 stage = getattr(m, defaultStage) 112 if name in stage.event: 113 if not name in events: 114 events[name] = [] 115 try: 116 events[name].append((max(stage.event[name].Time), sum(stage.event[name].Flops)/(max(stage.event[name].Time) * 1e6))) 117 except ZeroDivisionError: 118 events[name].append((max(stage.event[name].Time), 0)) 119 return 120 121def plotTime(library, num, eventNames, sizes, times, events): 122 from pylab import legend, plot, show, title, xlabel, ylabel 123 import numpy as np 124 125 arches = sizes.keys() 126 data = [] 127 for arch in arches: 128 data.append(sizes[arch]) 129 data.append(times[arch]) 130 plot(*data) 131 title('Performance on '+library+' Example '+str(num)) 132 xlabel('Number of Dof') 133 ylabel('Time (s)') 134 legend(arches, 'upper left', shadow = True) 135 show() 136 return 137 138def plotEventTime(library, num, eventNames, sizes, times, events, filename = None): 139 from pylab import close, legend, plot, savefig, show, title, xlabel, ylabel 140 import numpy as np 141 142 close() 143 arches = sizes.keys() 144 bs = events[arches[0]].keys()[0] 145 data = [] 146 names = [] 147 for event, color in zip(eventNames, ['b', 'g', 'r', 'y']): 148 for arch, style in zip(arches, ['-', ':']): 149 if event in events[arch][bs]: 150 names.append(arch+'-'+str(bs)+' '+event) 151 data.append(sizes[arch][bs]) 152 data.append(np.array(events[arch][bs][event])[:,0]) 153 data.append(color+style) 154 else: 155 print 'Could not find %s in %s-%d events' % (event, arch, bs) 156 print data 157 plot(*data) 158 title('Performance on '+library+' Example '+str(num)) 159 xlabel('Number of Dof') 160 ylabel('Time (s)') 161 legend(names, 'upper left', shadow = True) 162 if filename is None: 163 show() 164 else: 165 savefig(filename) 166 return 167 168def plotEventFlop(library, num, eventNames, sizes, times, events, filename = None): 169 from pylab import legend, plot, savefig, semilogy, show, title, xlabel, ylabel 170 import numpy as np 171 172 arches = sizes.keys() 173 bs = events[arches[0]].keys()[0] 174 data = [] 175 names = [] 176 for event, color in zip(eventNames, ['b', 'g', 'r', 'y']): 177 for arch, style in zip(arches, ['-', ':']): 178 if event in events[arch][bs]: 179 names.append(arch+'-'+str(bs)+' '+event) 180 data.append(sizes[arch][bs]) 181 data.append(1e-3*np.array(events[arch][bs][event])[:,1]) 182 data.append(color+style) 183 else: 184 print 'Could not find %s in %s-%d events' % (event, arch, bs) 185 semilogy(*data) 186 title('Performance on '+library+' Example '+str(num)) 187 xlabel('Number of Dof') 188 ylabel('Computation Rate (GF/s)') 189 legend(names, 'upper left', shadow = True) 190 if filename is None: 191 show() 192 else: 193 savefig(filename) 194 return 195 196def plotEventScaling(library, num, eventNames, procs, events, filename = None): 197 from pylab import legend, plot, savefig, semilogy, show, title, xlabel, ylabel 198 import numpy as np 199 200 arches = procs.keys() 201 bs = events[arches[0]].keys()[0] 202 data = [] 203 names = [] 204 for arch, style in zip(arches, ['-', ':']): 205 for event, color in zip(eventNames, ['b', 'g', 'r', 'y']): 206 if event in events[arch][bs]: 207 names.append(arch+'-'+str(bs)+' '+event) 208 data.append(procs[arch][bs]) 209 data.append(1e-3*np.array(events[arch][bs][event])[:,1]) 210 data.append(color+style) 211 else: 212 print 'Could not find %s in %s-%d events' % (event, arch, bs) 213 plot(*data) 214 title('Performance on '+library+' Example '+str(num)) 215 xlabel('Number of Processors') 216 ylabel('Computation Rate (GF/s)') 217 legend(names, 'upper left', shadow = True) 218 if filename is None: 219 show() 220 else: 221 savefig(filename) 222 return 223 224def plotSummaryLine(library, num, eventNames, sizes, times, events): 225 from pylab import legend, plot, show, title, xlabel, ylabel 226 import numpy as np 227 showTime = False 228 showEventTime = True 229 showEventFlops = True 230 arches = sizes.keys() 231 # Time 232 if showTime: 233 data = [] 234 for arch in arches: 235 data.append(sizes[arch]) 236 data.append(times[arch]) 237 plot(*data) 238 title('Performance on '+library+' Example '+str(num)) 239 xlabel('Number of Dof') 240 ylabel('Time (s)') 241 legend(arches, 'upper left', shadow = True) 242 show() 243 # Common event time 244 # We could make a stacked plot like Rio uses here 245 if showEventTime: 246 bs = events[arches[0]].keys()[0] 247 data = [] 248 names = [] 249 for event, color in zip(eventNames, ['b', 'g', 'r', 'y']): 250 for arch, style in zip(arches, ['-', ':']): 251 if event in events[arch][bs]: 252 names.append(arch+'-'+str(bs)+' '+event) 253 data.append(sizes[arch][bs]) 254 data.append(np.array(events[arch][bs][event])[:,0]) 255 data.append(color+style) 256 else: 257 print 'Could not find %s in %s-%d events' % (event, arch, bs) 258 print data 259 plot(*data) 260 title('Performance on '+library+' Example '+str(num)) 261 xlabel('Number of Dof') 262 ylabel('Time (s)') 263 legend(names, 'upper left', shadow = True) 264 show() 265 # Common event flops 266 # We could make a stacked plot like Rio uses here 267 if showEventFlops: 268 bs = events[arches[0]].keys()[0] 269 data = [] 270 names = [] 271 for event, color in zip(eventNames, ['b', 'g', 'r', 'y']): 272 for arch, style in zip(arches, ['-', ':']): 273 if event in events[arch][bs]: 274 names.append(arch+'-'+str(bs)+' '+event) 275 data.append(sizes[arch][bs]) 276 data.append(np.array(events[arch][bs][event])[:,1]) 277 data.append(color+style) 278 else: 279 print 'Could not find %s in %s-%d events' % (event, arch, bs) 280 plot(*data) 281 title('Performance on '+library+' Example '+str(num)) 282 xlabel('Number of Dof') 283 ylabel('Computation Rate (MF/s)') 284 legend(names, 'upper left', shadow = True) 285 show() 286 return 287 288def plotSummaryBar(library, num, eventNames, sizes, times, events): 289 import numpy as np 290 import matplotlib.pyplot as plt 291 292 eventColors = ['b', 'g', 'r', 'y'] 293 arches = sizes.keys() 294 names = [] 295 N = len(sizes[arches[0]]) 296 width = 0.2 297 ind = np.arange(N) - 0.25 298 bars = {} 299 for arch in arches: 300 bars[arch] = [] 301 bottom = np.zeros(N) 302 for event, color in zip(eventNames, eventColors): 303 names.append(arch+' '+event) 304 times = np.array(events[arch][event])[:,0] 305 bars[arch].append(plt.bar(ind, times, width, color=color, bottom=bottom)) 306 bottom += times 307 ind += 0.3 308 309 plt.xlabel('Number of Dof') 310 plt.ylabel('Time (s)') 311 plt.title('GPU vs. CPU Performance on '+library+' Example '+str(num)) 312 plt.xticks(np.arange(N), map(str, sizes[arches[0]])) 313 #plt.yticks(np.arange(0,81,10)) 314 #plt.legend( (p1[0], p2[0]), ('Men', 'Women') ) 315 plt.legend([bar[0] for bar in bars[arches[0]]], eventNames, 'upper right', shadow = True) 316 317 plt.show() 318 return 319 320def getDMComplexSize(dim, out): 321 '''Retrieves the number of cells from -dm_view output''' 322 size = 0 323 for line in out.split('\n'): 324 if line.strip().startswith(str(dim)+'-cells: '): 325 sizes = line.strip()[9:].split() 326 size = sum(map(int, sizes)) 327 break 328 return size 329 330def run_DMDA(ex, name, opts, args, sizes, times, events, log=True): 331 for n in map(int, args.size): 332 ex.run(log=log, da_grid_x=n, da_grid_y=n, **opts) 333 sizes[name].append(n*n * args.comp) 334 processSummary('summary', args.stage, args.events, times[name], events[name]) 335 return 336 337def run_DMComplex(ex, name, opts, args, sizes, times, events, log=True): 338 # This should eventually be replaced by a direct FFC/Ignition interface 339 if args.operator == 'laplacian': 340 numComp = 1 341 elif args.operator == 'elasticity': 342 numComp = args.dim 343 else: 344 raise RuntimeError('Unknown operator: %s' % args.operator) 345 346 for numBlock in [2**i for i in map(int, args.blockExp)]: 347 opts['gpu_blocks'] = numBlock 348 args.files = ['['+','.join(source)+']'] 349 buildExample(args) 350 sizes[name][numBlock] = [] 351 times[name][numBlock] = [] 352 events[name][numBlock] = {} 353 for r in map(float, args.refine): 354 out = ex.run(log=log, refinement_limit=r, **opts) 355 sizes[name][numBlock].append(getDMComplexSize(args.dim, out)) 356 processSummary('summary', args.stage, args.events, times[name][numBlock], events[name][numBlock]) 357 return 358 359def outputData(sizes, times, events, name = 'output.py'): 360 if os.path.exists(name): 361 base, ext = os.path.splitext(name) 362 num = 1 363 while os.path.exists(base+str(num)+ext): 364 num += 1 365 name = base+str(num)+ext 366 with file(name, 'w') as f: 367 f.write('#PETSC_ARCH='+os.environ['PETSC_ARCH']+' '+' '.join(sys.argv)+'\n') 368 f.write('sizes = '+repr(sizes)+'\n') 369 f.write('times = '+repr(times)+'\n') 370 f.write('events = '+repr(events)+'\n') 371 return 372 373if __name__ == '__main__': 374 import argparse 375 376 parser = argparse.ArgumentParser(description = 'PETSc Benchmarking', 377 epilog = 'This script runs src/<library>/examples/tutorials/ex<num>, For more information, visit http://www.mcs.anl.gov/petsc', 378 formatter_class = argparse.ArgumentDefaultsHelpFormatter) 379 parser.add_argument('--library', default='SNES', help='The PETSc library used in this example') 380 parser.add_argument('--num', type = int, default='5', help='The example number') 381 parser.add_argument('--module', default='summary', help='The module for timing output') 382 parser.add_argument('--stage', default='Main_Stage', help='The default logging stage') 383 parser.add_argument('--events', nargs='+', help='Events to process') 384 parser.add_argument('--batch', action='store_true', default=False, help='Generate batch files for the runs instead') 385 parser.add_argument('--daemon', action='store_true', default=False, help='Run as a daemon') 386 parser.add_argument('--gpulang', default='OpenCL', help='GPU Language to use: Either CUDA or OpenCL (default)') 387 subparsers = parser.add_subparsers(help='DM types') 388 389 parser_dmda = subparsers.add_parser('DMDA', help='Use a DMDA for the problem geometry') 390 parser_dmda.add_argument('--size', nargs='+', default=['10'], help='Grid size (implementation dependent)') 391 parser_dmda.add_argument('--comp', type = int, default='1', help='Number of field components') 392 parser_dmda.add_argument('runs', nargs='*', help='Run descriptions: <name>=<args>') 393 394 parser_dmmesh = subparsers.add_parser('DMComplex', help='Use a DMComplex for the problem geometry') 395 parser_dmmesh.add_argument('--dim', type = int, default='2', help='Spatial dimension') 396 parser_dmmesh.add_argument('--refine', nargs='+', default=['0.0'], help='List of refinement limits') 397 parser_dmmesh.add_argument('--order', type = int, default='1', help='Order of the finite element') 398 parser_dmmesh.add_argument('--operator', default='laplacian', help='The operator name') 399 parser_dmmesh.add_argument('--blockExp', nargs='+', default=range(0, 5), help='List of block exponents j, block size is 2^j') 400 parser_dmmesh.add_argument('runs', nargs='*', help='Run descriptions: <name>=<args>') 401 402 args = parser.parse_args() 403 print(args) 404 if hasattr(args, 'comp'): 405 args.dmType = 'DMDA' 406 else: 407 args.dmType = 'DMComplex' 408 409 ex = PETScExample(args.library, args.num, log_summary='summary.dat', log_summary_python = None if args.batch else args.module+'.py', preload='off') 410 if args.gpulang == 'CUDA': 411 source = ex.petsc.source(args.library, args.num, '.cu') 412 else: 413 source = ex.petsc.source(args.library, args.num, 'OpenCL.c') # Using the convention of OpenCL code residing in source files ending in 'OpenCL.c' (at least for snes/ex52) 414 sizes = {} 415 times = {} 416 events = {} 417 log = not args.daemon 418 419 if args.daemon: 420 import daemon 421 print 'Starting daemon' 422 daemon.createDaemon('.') 423 424 for run in args.runs: 425 name, stropts = run.split('=', 1) 426 opts = dict([t if len(t) == 2 else (t[0], None) for t in [arg.split('=', 1) for arg in stropts.split(' ')]]) 427 if args.dmType == 'DMDA': 428 sizes[name] = [] 429 times[name] = [] 430 events[name] = {} 431 run_DMDA(ex, name, opts, args, sizes, times, events, log=log) 432 elif args.dmType == 'DMComplex': 433 sizes[name] = {} 434 times[name] = {} 435 events[name] = {} 436 run_DMComplex(ex, name, opts, args, sizes, times, events, log=log) 437 outputData(sizes, times, events) 438 if not args.batch and log: plotSummaryLine(args.library, args.num, args.events, sizes, times, events) 439# Benchmark for ex50 440# ./src/benchmarks/benchmarkExample.py --events VecMDot VecMAXPY KSPGMRESOrthog MatMult VecCUSPCopyTo VecCUSPCopyFrom MatCUSPCopyTo --num 50 DMDA --size 10 20 50 100 --comp 4 CPU='pc_type=none mat_no_inode dm_vec_type=seq dm_mat_type=seqaij' GPU='pc_type=none mat_no_inode dm_vec_type=seqcusp dm_mat_type=seqaijcusp cusp_synchronize' 441# Benchmark for ex52 442# ./src/benchmarks/benchmarkExample.py --events IntegBatchCPU IntegBatchGPU IntegGPUOnly --num 52 DMComplex --refine 0.0625 0.00625 0.000625 0.0000625 --blockExp 4 --order=1 CPU='dm_view show_residual=0 compute_function batch' GPU='dm_view show_residual=0 compute_function batch gpu gpu_batches=8' 443# ./src/benchmarks/benchmarkExample.py --events IntegBatchCPU IntegBatchGPU IntegGPUOnly --num 52 DMComplex --refine 0.0625 0.00625 0.000625 0.0000625 --blockExp 4 --order=1 --operator=elasticity CPU='dm_view op_type=elasticity show_residual=0 compute_function batch' GPU='dm_view op_type=elasticity show_residual=0 compute_function batch gpu gpu_batches=8' 444# ./src/benchmarks/benchmarkExample.py --events IntegBatchCPU IntegBatchGPU IntegGPUOnly --num 52 DMComplex --dim=3 --refine 0.0625 0.00625 0.000625 0.0000625 --blockExp 4 --order=1 CPU='dim=3 dm_view show_residual=0 compute_function batch' GPU='dim=3 dm_view show_residual=0 compute_function batch gpu gpu_batches=8' 445