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