| 1 | #!/usr/bin/env python |
|---|
| 2 | |
|---|
| 3 | # Specify debugging level here; |
|---|
| 4 | # |
|---|
| 5 | # 10 = gemtric cmd's |
|---|
| 6 | DEBUG_LEVEL = 0 |
|---|
| 7 | |
|---|
| 8 | # Wether or not to run as a daemon in background |
|---|
| 9 | # |
|---|
| 10 | DAEMONIZE = 1 |
|---|
| 11 | |
|---|
| 12 | # Which Torque server to monitor |
|---|
| 13 | # |
|---|
| 14 | TORQUE_SERVER = 'localhost' |
|---|
| 15 | |
|---|
| 16 | # How many seconds interval for polling of jobs |
|---|
| 17 | # |
|---|
| 18 | # this will effect directly how accurate the |
|---|
| 19 | # end time of a job can be determined |
|---|
| 20 | # |
|---|
| 21 | TORQUE_POLL_INTERVAL = 10 |
|---|
| 22 | |
|---|
| 23 | # Alternate location of gmond.conf |
|---|
| 24 | # |
|---|
| 25 | # Default: /etc/gmond.conf |
|---|
| 26 | # |
|---|
| 27 | #GMOND_CONF = '/etc/gmond.conf' |
|---|
| 28 | |
|---|
| 29 | from PBSQuery import PBSQuery |
|---|
| 30 | import sys |
|---|
| 31 | import time |
|---|
| 32 | import os |
|---|
| 33 | import socket |
|---|
| 34 | import string |
|---|
| 35 | |
|---|
| 36 | class DataProcessor: |
|---|
| 37 | """Class for processing of data""" |
|---|
| 38 | |
|---|
| 39 | binary = '/usr/bin/gmetric' |
|---|
| 40 | |
|---|
| 41 | def __init__( self, binary=None ): |
|---|
| 42 | """Remember alternate binary location if supplied""" |
|---|
| 43 | |
|---|
| 44 | if binary: |
|---|
| 45 | self.binary = binary |
|---|
| 46 | |
|---|
| 47 | # Timeout for XML |
|---|
| 48 | # |
|---|
| 49 | # From ganglia's documentation: |
|---|
| 50 | # |
|---|
| 51 | # 'A metric will be deleted DMAX seconds after it is received, and |
|---|
| 52 | # DMAX=0 means eternal life.' |
|---|
| 53 | |
|---|
| 54 | self.dmax = str( int( TORQUE_POLL_INTERVAL ) ) |
|---|
| 55 | |
|---|
| 56 | try: |
|---|
| 57 | gmond_file = GMOND_CONF |
|---|
| 58 | |
|---|
| 59 | except NameError: |
|---|
| 60 | gmond_file = '/etc/gmond.conf' |
|---|
| 61 | |
|---|
| 62 | if not os.path.exists( gmond_file ): |
|---|
| 63 | debug_msg( 0, gmond_file + ' does not exist' ) |
|---|
| 64 | sys.exit( 1 ) |
|---|
| 65 | |
|---|
| 66 | incompatible = self.checkGmetricVersion() |
|---|
| 67 | |
|---|
| 68 | if incompatible: |
|---|
| 69 | debug_msg( 0, 'Gmetric version not compatible, pls upgrade to at least 3.0.1' ) |
|---|
| 70 | sys.exit( 1 ) |
|---|
| 71 | |
|---|
| 72 | def checkGmetricVersion( self ): |
|---|
| 73 | """ |
|---|
| 74 | Check version of gmetric is at least 3.0.1 |
|---|
| 75 | for the syntax we use |
|---|
| 76 | """ |
|---|
| 77 | |
|---|
| 78 | for line in os.popen( self.binary + ' --version' ).readlines(): |
|---|
| 79 | |
|---|
| 80 | line = line.split( ' ' ) |
|---|
| 81 | |
|---|
| 82 | if len( line ) == 2 and str(line).find( 'gmetric' ) != -1: |
|---|
| 83 | |
|---|
| 84 | gmetric_version = line[1].split( '\n' )[0] |
|---|
| 85 | |
|---|
| 86 | version_major = int( gmetric_version.split( '.' )[0] ) |
|---|
| 87 | version_minor = int( gmetric_version.split( '.' )[1] ) |
|---|
| 88 | version_patch = int( gmetric_version.split( '.' )[2] ) |
|---|
| 89 | |
|---|
| 90 | incompatible = 0 |
|---|
| 91 | |
|---|
| 92 | if version_major < 3: |
|---|
| 93 | |
|---|
| 94 | incompatible = 1 |
|---|
| 95 | |
|---|
| 96 | elif version_major == 3: |
|---|
| 97 | |
|---|
| 98 | if version_minor == 0: |
|---|
| 99 | |
|---|
| 100 | if version_patch < 1: |
|---|
| 101 | |
|---|
| 102 | incompatible = 1 |
|---|
| 103 | |
|---|
| 104 | return incompatible |
|---|
| 105 | |
|---|
| 106 | def multicastGmetric( self, metricname, metricval, valtype='string' ): |
|---|
| 107 | """Call gmetric binary and multicast""" |
|---|
| 108 | |
|---|
| 109 | cmd = self.binary |
|---|
| 110 | |
|---|
| 111 | try: |
|---|
| 112 | cmd = cmd + ' -c' + GMOND_CONF |
|---|
| 113 | except NameError: |
|---|
| 114 | debug_msg( 10, 'Assuming /etc/gmond.conf for gmetric cmd (ommitting)' ) |
|---|
| 115 | |
|---|
| 116 | cmd = cmd + ' -n' + str( metricname )+ ' -v"' + str( metricval )+ '" -t' + str( valtype ) + ' -d' + str( self.dmax ) |
|---|
| 117 | |
|---|
| 118 | debug_msg( 10, printTime() + ' ' + cmd ) |
|---|
| 119 | os.system( cmd ) |
|---|
| 120 | |
|---|
| 121 | class PBSDataGatherer: |
|---|
| 122 | |
|---|
| 123 | jobs = { } |
|---|
| 124 | |
|---|
| 125 | def __init__( self ): |
|---|
| 126 | """Setup appropriate variables""" |
|---|
| 127 | |
|---|
| 128 | self.jobs = { } |
|---|
| 129 | self.dp = DataProcessor() |
|---|
| 130 | self.initPbsQuery() |
|---|
| 131 | |
|---|
| 132 | def initPbsQuery( self ): |
|---|
| 133 | |
|---|
| 134 | self.pq = None |
|---|
| 135 | if( TORQUE_SERVER ): |
|---|
| 136 | self.pq = PBSQuery( TORQUE_SERVER ) |
|---|
| 137 | else |
|---|
| 138 | self.pq = PBSQuery() |
|---|
| 139 | |
|---|
| 140 | def getAttr( self, attrs, name ): |
|---|
| 141 | """Return certain attribute from dictionary, if exists""" |
|---|
| 142 | |
|---|
| 143 | if attrs.has_key( name ): |
|---|
| 144 | return attrs[name] |
|---|
| 145 | else: |
|---|
| 146 | return '' |
|---|
| 147 | |
|---|
| 148 | def jobDataChanged( self, jobs, job_id, attrs ): |
|---|
| 149 | """Check if job with attrs and job_id in jobs has changed""" |
|---|
| 150 | |
|---|
| 151 | if jobs.has_key( job_id ): |
|---|
| 152 | oldData = jobs[ job_id ] |
|---|
| 153 | else: |
|---|
| 154 | return 1 |
|---|
| 155 | |
|---|
| 156 | for name, val in attrs.items(): |
|---|
| 157 | |
|---|
| 158 | if oldData.has_key( name ): |
|---|
| 159 | |
|---|
| 160 | if oldData[ name ] != attrs[ name ]: |
|---|
| 161 | |
|---|
| 162 | return 1 |
|---|
| 163 | |
|---|
| 164 | else: |
|---|
| 165 | return 1 |
|---|
| 166 | |
|---|
| 167 | return 0 |
|---|
| 168 | |
|---|
| 169 | def getJobData( self, known_jobs ): |
|---|
| 170 | """Gather all data on current jobs in Torque""" |
|---|
| 171 | |
|---|
| 172 | if len( known_jobs ) > 0: |
|---|
| 173 | jobs = known_jobs |
|---|
| 174 | else: |
|---|
| 175 | jobs = { } |
|---|
| 176 | |
|---|
| 177 | #self.initPbsQuery() |
|---|
| 178 | |
|---|
| 179 | #print self.pq.getnodes() |
|---|
| 180 | |
|---|
| 181 | joblist = self.pq.getjobs() |
|---|
| 182 | |
|---|
| 183 | self.cur_time = time.time() |
|---|
| 184 | |
|---|
| 185 | jobs_processed = [ ] |
|---|
| 186 | |
|---|
| 187 | #self.printJobs( joblist ) |
|---|
| 188 | |
|---|
| 189 | for name, attrs in joblist.items(): |
|---|
| 190 | |
|---|
| 191 | job_id = name.split( '.' )[0] |
|---|
| 192 | |
|---|
| 193 | jobs_processed.append( job_id ) |
|---|
| 194 | |
|---|
| 195 | name = self.getAttr( attrs, 'Job_Name' ) |
|---|
| 196 | queue = self.getAttr( attrs, 'queue' ) |
|---|
| 197 | owner = self.getAttr( attrs, 'Job_Owner' ).split( '@' )[0] |
|---|
| 198 | requested_time = self.getAttr( attrs, 'Resource_List.walltime' ) |
|---|
| 199 | requested_memory = self.getAttr( attrs, 'Resource_List.mem' ) |
|---|
| 200 | |
|---|
| 201 | mynoderequest = self.getAttr( attrs, 'Resource_List.nodes' ) |
|---|
| 202 | |
|---|
| 203 | if mynoderequest.find( ':' ) != -1 and mynoderequest.find( 'ppn' ) != -1: |
|---|
| 204 | ppn = mynoderequest.split( ':' )[1].split( 'ppn=' )[1] |
|---|
| 205 | else: |
|---|
| 206 | ppn = '' |
|---|
| 207 | |
|---|
| 208 | status = self.getAttr( attrs, 'job_state' ) |
|---|
| 209 | |
|---|
| 210 | if status == 'R': |
|---|
| 211 | start_timestamp = self.getAttr( attrs, 'mtime' ) |
|---|
| 212 | nodes = self.getAttr( attrs, 'exec_host' ).split( '+' ) |
|---|
| 213 | |
|---|
| 214 | nodeslist = [ ] |
|---|
| 215 | |
|---|
| 216 | for node in nodes: |
|---|
| 217 | host = node.split( '/' )[0] |
|---|
| 218 | |
|---|
| 219 | if nodeslist.count( host ) == 0: |
|---|
| 220 | nodeslist.append( host ) |
|---|
| 221 | |
|---|
| 222 | elif status == 'Q': |
|---|
| 223 | start_timestamp = '' |
|---|
| 224 | count_mynodes = 0 |
|---|
| 225 | numeric_node = 1 |
|---|
| 226 | |
|---|
| 227 | for node in mynoderequest.split( '+' ): |
|---|
| 228 | |
|---|
| 229 | nodepart = node.split( ':' )[0] |
|---|
| 230 | |
|---|
| 231 | for letter in nodepart: |
|---|
| 232 | |
|---|
| 233 | if letter not in string.digits: |
|---|
| 234 | |
|---|
| 235 | numeric_node = 0 |
|---|
| 236 | |
|---|
| 237 | if not numeric_node: |
|---|
| 238 | count_mynodes = count_mynodes + 1 |
|---|
| 239 | else: |
|---|
| 240 | count_mynodes = count_mynodes + int( nodepart ) |
|---|
| 241 | |
|---|
| 242 | nodeslist = count_mynodes |
|---|
| 243 | else: |
|---|
| 244 | start_timestamp = '' |
|---|
| 245 | nodeslit = '' |
|---|
| 246 | |
|---|
| 247 | myAttrs = { } |
|---|
| 248 | myAttrs['name'] = name |
|---|
| 249 | myAttrs['queue'] = queue |
|---|
| 250 | myAttrs['owner'] = owner |
|---|
| 251 | myAttrs['requested_time'] = requested_time |
|---|
| 252 | myAttrs['requested_memory'] = requested_memory |
|---|
| 253 | myAttrs['ppn'] = ppn |
|---|
| 254 | myAttrs['status'] = status |
|---|
| 255 | myAttrs['start_timestamp'] = start_timestamp |
|---|
| 256 | myAttrs['reported'] = str( int( self.cur_time ) ) |
|---|
| 257 | myAttrs['nodes'] = nodeslist |
|---|
| 258 | myAttrs['domain'] = string.join( socket.getfqdn().split( '.' )[1:], '.' ) |
|---|
| 259 | myAttrs['poll_interval'] = TORQUE_POLL_INTERVAL |
|---|
| 260 | |
|---|
| 261 | if self.jobDataChanged( jobs, job_id, myAttrs ): |
|---|
| 262 | jobs[ job_id ] = myAttrs |
|---|
| 263 | |
|---|
| 264 | #debug_msg( 10, printTime() + ' job %s state changed' %(job_id) ) |
|---|
| 265 | |
|---|
| 266 | for id, attrs in jobs.items(): |
|---|
| 267 | |
|---|
| 268 | if id not in jobs_processed: |
|---|
| 269 | |
|---|
| 270 | # This one isn't there anymore; toedeledoki! |
|---|
| 271 | # |
|---|
| 272 | del jobs[ id ] |
|---|
| 273 | |
|---|
| 274 | return jobs |
|---|
| 275 | |
|---|
| 276 | def submitJobData( self, jobs ): |
|---|
| 277 | """Submit job info list""" |
|---|
| 278 | |
|---|
| 279 | self.dp.multicastGmetric( 'TOGA-HEARTBEAT', str( int( self.cur_time ) ) ) |
|---|
| 280 | |
|---|
| 281 | # Now let's spread the knowledge |
|---|
| 282 | # |
|---|
| 283 | for jobid, jobattrs in jobs.items(): |
|---|
| 284 | |
|---|
| 285 | gmetric_val = self.compileGmetricVal( jobid, jobattrs ) |
|---|
| 286 | |
|---|
| 287 | for val in gmetric_val: |
|---|
| 288 | self.dp.multicastGmetric( 'TOGA-JOB-' + jobid, val ) |
|---|
| 289 | |
|---|
| 290 | def makeNodeString( self, nodelist ): |
|---|
| 291 | """Make one big string of all hosts""" |
|---|
| 292 | |
|---|
| 293 | node_str = None |
|---|
| 294 | |
|---|
| 295 | for node in nodelist: |
|---|
| 296 | if not node_str: |
|---|
| 297 | node_str = node |
|---|
| 298 | else: |
|---|
| 299 | node_str = node_str + ';' + node |
|---|
| 300 | |
|---|
| 301 | return node_str |
|---|
| 302 | |
|---|
| 303 | def compileGmetricVal( self, jobid, jobattrs ): |
|---|
| 304 | """Create a val string for gmetric of jobinfo""" |
|---|
| 305 | |
|---|
| 306 | appendList = [ ] |
|---|
| 307 | appendList.append( 'name=' + jobattrs['name'] ) |
|---|
| 308 | appendList.append( 'queue=' + jobattrs['queue'] ) |
|---|
| 309 | appendList.append( 'owner=' + jobattrs['owner'] ) |
|---|
| 310 | appendList.append( 'requested_time=' + jobattrs['requested_time'] ) |
|---|
| 311 | |
|---|
| 312 | if jobattrs['requested_memory'] != '': |
|---|
| 313 | appendList.append( 'requested_memory=' + jobattrs['requested_memory'] ) |
|---|
| 314 | |
|---|
| 315 | if jobattrs['ppn'] != '': |
|---|
| 316 | appendList.append( 'ppn=' + jobattrs['ppn'] ) |
|---|
| 317 | |
|---|
| 318 | appendList.append( 'status=' + jobattrs['status'] ) |
|---|
| 319 | |
|---|
| 320 | if jobattrs['start_timestamp'] != '': |
|---|
| 321 | appendList.append( 'start_timestamp=' + jobattrs['start_timestamp'] ) |
|---|
| 322 | |
|---|
| 323 | appendList.append( 'reported=' + jobattrs['reported'] ) |
|---|
| 324 | appendList.append( 'poll_interval=' + str( jobattrs['poll_interval'] ) ) |
|---|
| 325 | appendList.append( 'domain=' + jobattrs['domain'] ) |
|---|
| 326 | |
|---|
| 327 | if jobattrs['status'] == 'R': |
|---|
| 328 | if len( jobattrs['nodes'] ) > 0: |
|---|
| 329 | appendList.append( 'nodes=' + self.makeNodeString( jobattrs['nodes'] ) ) |
|---|
| 330 | elif jobattrs['status'] == 'Q': |
|---|
| 331 | appendList.append( 'nodes=' + str(jobattrs['nodes']) ) |
|---|
| 332 | |
|---|
| 333 | return self.makeAppendLists( appendList ) |
|---|
| 334 | |
|---|
| 335 | def makeAppendLists( self, append_list ): |
|---|
| 336 | """ |
|---|
| 337 | Divide all values from append_list over strings with a maximum |
|---|
| 338 | size of 1400 |
|---|
| 339 | """ |
|---|
| 340 | |
|---|
| 341 | app_lists = [ ] |
|---|
| 342 | |
|---|
| 343 | mystr = None |
|---|
| 344 | |
|---|
| 345 | for val in append_list: |
|---|
| 346 | |
|---|
| 347 | if not mystr: |
|---|
| 348 | mystr = val |
|---|
| 349 | else: |
|---|
| 350 | if not self.checkValAppendMaxSize( mystr, val ): |
|---|
| 351 | mystr = mystr + ' ' + val |
|---|
| 352 | else: |
|---|
| 353 | # Too big, new appenlist |
|---|
| 354 | app_lists.append( mystr ) |
|---|
| 355 | mystr = val |
|---|
| 356 | |
|---|
| 357 | app_lists.append( mystr ) |
|---|
| 358 | |
|---|
| 359 | return app_lists |
|---|
| 360 | |
|---|
| 361 | def checkValAppendMaxSize( self, val, text ): |
|---|
| 362 | """Check if val + text size is not above 1400 (max msg size)""" |
|---|
| 363 | |
|---|
| 364 | # Max frame size of a udp datagram is 1500 bytes |
|---|
| 365 | # removing misc header and gmetric stuff leaves about 1400 bytes |
|---|
| 366 | # |
|---|
| 367 | if len( val + text ) > 1400: |
|---|
| 368 | return 1 |
|---|
| 369 | else: |
|---|
| 370 | return 0 |
|---|
| 371 | |
|---|
| 372 | def printJobs( self, jobs ): |
|---|
| 373 | """Print a jobinfo overview""" |
|---|
| 374 | |
|---|
| 375 | for name, attrs in self.jobs.items(): |
|---|
| 376 | |
|---|
| 377 | print 'job %s' %(name) |
|---|
| 378 | |
|---|
| 379 | for name, val in attrs.items(): |
|---|
| 380 | |
|---|
| 381 | print '\t%s = %s' %( name, val ) |
|---|
| 382 | |
|---|
| 383 | def printJob( self, jobs, job_id ): |
|---|
| 384 | """Print job with job_id from jobs""" |
|---|
| 385 | |
|---|
| 386 | print 'job %s' %(job_id) |
|---|
| 387 | |
|---|
| 388 | for name, val in jobs[ job_id ].items(): |
|---|
| 389 | |
|---|
| 390 | print '\t%s = %s' %( name, val ) |
|---|
| 391 | |
|---|
| 392 | def daemon( self ): |
|---|
| 393 | """Run as daemon forever""" |
|---|
| 394 | |
|---|
| 395 | # Fork the first child |
|---|
| 396 | # |
|---|
| 397 | pid = os.fork() |
|---|
| 398 | if pid > 0: |
|---|
| 399 | sys.exit(0) # end parrent |
|---|
| 400 | |
|---|
| 401 | # creates a session and sets the process group ID |
|---|
| 402 | # |
|---|
| 403 | os.setsid() |
|---|
| 404 | |
|---|
| 405 | # Fork the second child |
|---|
| 406 | # |
|---|
| 407 | pid = os.fork() |
|---|
| 408 | if pid > 0: |
|---|
| 409 | sys.exit(0) # end parrent |
|---|
| 410 | |
|---|
| 411 | # Go to the root directory and set the umask |
|---|
| 412 | # |
|---|
| 413 | os.chdir('/') |
|---|
| 414 | os.umask(0) |
|---|
| 415 | |
|---|
| 416 | sys.stdin.close() |
|---|
| 417 | sys.stdout.close() |
|---|
| 418 | sys.stderr.close() |
|---|
| 419 | |
|---|
| 420 | os.open('/dev/null', 0) |
|---|
| 421 | os.dup(0) |
|---|
| 422 | os.dup(0) |
|---|
| 423 | |
|---|
| 424 | self.run() |
|---|
| 425 | |
|---|
| 426 | def run( self ): |
|---|
| 427 | """Main thread""" |
|---|
| 428 | |
|---|
| 429 | while ( 1 ): |
|---|
| 430 | |
|---|
| 431 | self.jobs = self.getJobData( self.jobs ) |
|---|
| 432 | self.submitJobData( self.jobs ) |
|---|
| 433 | time.sleep( TORQUE_POLL_INTERVAL ) |
|---|
| 434 | |
|---|
| 435 | def printTime( ): |
|---|
| 436 | """Print current time/date in human readable format for log/debug""" |
|---|
| 437 | |
|---|
| 438 | return time.strftime("%a, %d %b %Y %H:%M:%S") |
|---|
| 439 | |
|---|
| 440 | def debug_msg( level, msg ): |
|---|
| 441 | """Print msg if at or above current debug level""" |
|---|
| 442 | |
|---|
| 443 | if (DEBUG_LEVEL >= level): |
|---|
| 444 | sys.stderr.write( msg + '\n' ) |
|---|
| 445 | |
|---|
| 446 | def main(): |
|---|
| 447 | """Application start""" |
|---|
| 448 | |
|---|
| 449 | gather = PBSDataGatherer() |
|---|
| 450 | if DAEMONIZE: |
|---|
| 451 | gather.daemon() |
|---|
| 452 | else: |
|---|
| 453 | gather.run() |
|---|
| 454 | |
|---|
| 455 | # w00t someone started me |
|---|
| 456 | # |
|---|
| 457 | if __name__ == '__main__': |
|---|
| 458 | main() |
|---|
