1 | #!/usr/bin/env python |
---|
2 | |
---|
3 | import sys, getopt, ConfigParser |
---|
4 | |
---|
5 | def processArgs( args ): |
---|
6 | |
---|
7 | SHORT_L = 'c:' |
---|
8 | LONG_L = 'config=' |
---|
9 | |
---|
10 | config_filename = None |
---|
11 | |
---|
12 | try: |
---|
13 | |
---|
14 | opts, args = getopt.getopt( args, SHORT_L, LONG_L ) |
---|
15 | |
---|
16 | except getopt.error, detail: |
---|
17 | |
---|
18 | print detail |
---|
19 | sys.exit(1) |
---|
20 | |
---|
21 | for opt, value in opts: |
---|
22 | |
---|
23 | if opt in [ '--config', '-c' ]: |
---|
24 | |
---|
25 | config_filename = value |
---|
26 | |
---|
27 | if not config_filename: |
---|
28 | |
---|
29 | config_filename = '/etc/jobmond.conf' |
---|
30 | |
---|
31 | return loadConfig( config_filename ) |
---|
32 | |
---|
33 | def loadConfig( filename ): |
---|
34 | |
---|
35 | def getlist( cfg_string ): |
---|
36 | |
---|
37 | my_list = [ ] |
---|
38 | |
---|
39 | for item_txt in cfg_string.split( ',' ): |
---|
40 | |
---|
41 | sep_char = None |
---|
42 | |
---|
43 | item_txt = item_txt.strip() |
---|
44 | |
---|
45 | for s_char in [ "'", '"' ]: |
---|
46 | |
---|
47 | if item_txt.find( s_char ) != -1: |
---|
48 | |
---|
49 | if item_txt.count( s_char ) != 2: |
---|
50 | |
---|
51 | print 'Missing quote: %s' %item_txt |
---|
52 | sys.exit( 1 ) |
---|
53 | |
---|
54 | else: |
---|
55 | |
---|
56 | sep_char = s_char |
---|
57 | break |
---|
58 | |
---|
59 | if sep_char: |
---|
60 | |
---|
61 | item_txt = item_txt.split( sep_char )[1] |
---|
62 | |
---|
63 | my_list.append( item_txt ) |
---|
64 | |
---|
65 | return my_list |
---|
66 | |
---|
67 | cfg = ConfigParser.ConfigParser() |
---|
68 | |
---|
69 | cfg.read( filename ) |
---|
70 | |
---|
71 | global DEBUG_LEVEL, DAEMONIZE, TORQUE_SERVER, TORQUE_POLL_INTERVAL, GMOND_CONF, DETECT_TIME_DIFFS, BATCH_HOST_TRANSLATE |
---|
72 | |
---|
73 | # Specify debugging level here; |
---|
74 | # |
---|
75 | # 10 = gemtric cmd's |
---|
76 | DEBUG_LEVEL = cfg.getint( 'DEFAULT', 'DEBUG_LEVEL' ) |
---|
77 | |
---|
78 | # Wether or not to run as a daemon in background |
---|
79 | # |
---|
80 | DAEMONIZE = cfg.getboolean( 'DEFAULT', 'DAEMONIZE' ) |
---|
81 | |
---|
82 | # Which Torque server to monitor |
---|
83 | # |
---|
84 | TORQUE_SERVER = cfg.get( 'DEFAULT', 'TORQUE_SERVER' ) |
---|
85 | |
---|
86 | # How many seconds interval for polling of jobs |
---|
87 | # |
---|
88 | # this will effect directly how accurate the |
---|
89 | # end time of a job can be determined |
---|
90 | # |
---|
91 | TORQUE_POLL_INTERVAL = cfg.getint( 'DEFAULT', 'TORQUE_POLL_INTERVAL' ) |
---|
92 | |
---|
93 | # Alternate location of gmond.conf |
---|
94 | # |
---|
95 | # Default: /etc/gmond.conf |
---|
96 | # |
---|
97 | GMOND_CONF = cfg.get( 'DEFAULT', 'GMOND_CONF' ) |
---|
98 | |
---|
99 | # Wether or not to detect differences in |
---|
100 | # time from Torque server and local time. |
---|
101 | # |
---|
102 | # Ideally both machines (if not the same) |
---|
103 | # should have the same time (via ntp or whatever) |
---|
104 | # |
---|
105 | DETECT_TIME_DIFFS = cfg.getboolean( 'DEFAULT', 'DETECT_TIME_DIFFS' ) |
---|
106 | |
---|
107 | # Regexp style hostname translation |
---|
108 | # |
---|
109 | # Usefull if your Torque hostnames are not the same as your |
---|
110 | # Ganglia hostnames (different network interfaces) |
---|
111 | # |
---|
112 | # Syntax: /orig/new/ |
---|
113 | # |
---|
114 | BATCH_HOST_TRANSLATE = getlist( cfg.get( 'DEFAULT', 'BATCH_HOST_TRANSLATE' ) ) |
---|
115 | |
---|
116 | return True |
---|
117 | |
---|
118 | from PBSQuery import PBSQuery |
---|
119 | |
---|
120 | import time, os, socket, string, re |
---|
121 | |
---|
122 | class DataProcessor: |
---|
123 | """Class for processing of data""" |
---|
124 | |
---|
125 | binary = '/usr/bin/gmetric' |
---|
126 | |
---|
127 | def __init__( self, binary=None ): |
---|
128 | """Remember alternate binary location if supplied""" |
---|
129 | |
---|
130 | if binary: |
---|
131 | self.binary = binary |
---|
132 | |
---|
133 | # Timeout for XML |
---|
134 | # |
---|
135 | # From ganglia's documentation: |
---|
136 | # |
---|
137 | # 'A metric will be deleted DMAX seconds after it is received, and |
---|
138 | # DMAX=0 means eternal life.' |
---|
139 | |
---|
140 | self.dmax = str( int( int( TORQUE_POLL_INTERVAL ) + 2 ) ) |
---|
141 | |
---|
142 | try: |
---|
143 | gmond_file = GMOND_CONF |
---|
144 | |
---|
145 | except NameError: |
---|
146 | gmond_file = '/etc/gmond.conf' |
---|
147 | |
---|
148 | if not os.path.exists( gmond_file ): |
---|
149 | debug_msg( 0, gmond_file + ' does not exist' ) |
---|
150 | sys.exit( 1 ) |
---|
151 | |
---|
152 | incompatible = self.checkGmetricVersion() |
---|
153 | |
---|
154 | if incompatible: |
---|
155 | debug_msg( 0, 'Gmetric version not compatible, pls upgrade to at least 3.0.1' ) |
---|
156 | sys.exit( 1 ) |
---|
157 | |
---|
158 | def checkGmetricVersion( self ): |
---|
159 | """ |
---|
160 | Check version of gmetric is at least 3.0.1 |
---|
161 | for the syntax we use |
---|
162 | """ |
---|
163 | |
---|
164 | for line in os.popen( self.binary + ' --version' ).readlines(): |
---|
165 | |
---|
166 | line = line.split( ' ' ) |
---|
167 | |
---|
168 | if len( line ) == 2 and str(line).find( 'gmetric' ) != -1: |
---|
169 | |
---|
170 | gmetric_version = line[1].split( '\n' )[0] |
---|
171 | |
---|
172 | version_major = int( gmetric_version.split( '.' )[0] ) |
---|
173 | version_minor = int( gmetric_version.split( '.' )[1] ) |
---|
174 | version_patch = int( gmetric_version.split( '.' )[2] ) |
---|
175 | |
---|
176 | incompatible = 0 |
---|
177 | |
---|
178 | if version_major < 3: |
---|
179 | |
---|
180 | incompatible = 1 |
---|
181 | |
---|
182 | elif version_major == 3: |
---|
183 | |
---|
184 | if version_minor == 0: |
---|
185 | |
---|
186 | if version_patch < 1: |
---|
187 | |
---|
188 | incompatible = 1 |
---|
189 | |
---|
190 | return incompatible |
---|
191 | |
---|
192 | def multicastGmetric( self, metricname, metricval, valtype='string' ): |
---|
193 | """Call gmetric binary and multicast""" |
---|
194 | |
---|
195 | cmd = self.binary |
---|
196 | |
---|
197 | try: |
---|
198 | cmd = cmd + ' -c' + GMOND_CONF |
---|
199 | except NameError: |
---|
200 | debug_msg( 10, 'Assuming /etc/gmond.conf for gmetric cmd (ommitting)' ) |
---|
201 | |
---|
202 | cmd = cmd + ' -n' + str( metricname )+ ' -v"' + str( metricval )+ '" -t' + str( valtype ) + ' -d' + str( self.dmax ) |
---|
203 | |
---|
204 | debug_msg( 10, printTime() + ' ' + cmd ) |
---|
205 | os.system( cmd ) |
---|
206 | |
---|
207 | class DataGatherer: |
---|
208 | |
---|
209 | jobs = { } |
---|
210 | |
---|
211 | def __init__( self ): |
---|
212 | """Setup appropriate variables""" |
---|
213 | |
---|
214 | self.jobs = { } |
---|
215 | self.timeoffset = 0 |
---|
216 | self.dp = DataProcessor() |
---|
217 | self.initPbsQuery() |
---|
218 | |
---|
219 | def initPbsQuery( self ): |
---|
220 | |
---|
221 | self.pq = None |
---|
222 | if( TORQUE_SERVER ): |
---|
223 | self.pq = PBSQuery( TORQUE_SERVER ) |
---|
224 | else: |
---|
225 | self.pq = PBSQuery() |
---|
226 | |
---|
227 | def getAttr( self, attrs, name ): |
---|
228 | """Return certain attribute from dictionary, if exists""" |
---|
229 | |
---|
230 | if attrs.has_key( name ): |
---|
231 | return attrs[name] |
---|
232 | else: |
---|
233 | return '' |
---|
234 | |
---|
235 | def jobDataChanged( self, jobs, job_id, attrs ): |
---|
236 | """Check if job with attrs and job_id in jobs has changed""" |
---|
237 | |
---|
238 | if jobs.has_key( job_id ): |
---|
239 | oldData = jobs[ job_id ] |
---|
240 | else: |
---|
241 | return 1 |
---|
242 | |
---|
243 | for name, val in attrs.items(): |
---|
244 | |
---|
245 | if oldData.has_key( name ): |
---|
246 | |
---|
247 | if oldData[ name ] != attrs[ name ]: |
---|
248 | |
---|
249 | return 1 |
---|
250 | |
---|
251 | else: |
---|
252 | return 1 |
---|
253 | |
---|
254 | return 0 |
---|
255 | |
---|
256 | def getJobData( self, known_jobs ): |
---|
257 | """Gather all data on current jobs in Torque""" |
---|
258 | |
---|
259 | if len( known_jobs ) > 0: |
---|
260 | jobs = known_jobs |
---|
261 | else: |
---|
262 | jobs = { } |
---|
263 | |
---|
264 | #self.initPbsQuery() |
---|
265 | |
---|
266 | #print self.pq.getnodes() |
---|
267 | |
---|
268 | joblist = self.pq.getjobs() |
---|
269 | |
---|
270 | self.cur_time = time.time() |
---|
271 | |
---|
272 | jobs_processed = [ ] |
---|
273 | |
---|
274 | #self.printJobs( joblist ) |
---|
275 | |
---|
276 | for name, attrs in joblist.items(): |
---|
277 | |
---|
278 | job_id = name.split( '.' )[0] |
---|
279 | |
---|
280 | jobs_processed.append( job_id ) |
---|
281 | |
---|
282 | name = self.getAttr( attrs, 'Job_Name' ) |
---|
283 | queue = self.getAttr( attrs, 'queue' ) |
---|
284 | owner = self.getAttr( attrs, 'Job_Owner' ).split( '@' )[0] |
---|
285 | requested_time = self.getAttr( attrs, 'Resource_List.walltime' ) |
---|
286 | requested_memory = self.getAttr( attrs, 'Resource_List.mem' ) |
---|
287 | |
---|
288 | mynoderequest = self.getAttr( attrs, 'Resource_List.nodes' ) |
---|
289 | |
---|
290 | if mynoderequest.find( ':' ) != -1 and mynoderequest.find( 'ppn' ) != -1: |
---|
291 | ppn = mynoderequest.split( ':' )[1].split( 'ppn=' )[1] |
---|
292 | else: |
---|
293 | ppn = '' |
---|
294 | |
---|
295 | status = self.getAttr( attrs, 'job_state' ) |
---|
296 | |
---|
297 | if status == 'R': |
---|
298 | start_timestamp = self.getAttr( attrs, 'mtime' ) |
---|
299 | nodes = self.getAttr( attrs, 'exec_host' ).split( '+' ) |
---|
300 | |
---|
301 | nodeslist = [ ] |
---|
302 | |
---|
303 | for node in nodes: |
---|
304 | host = node.split( '/' )[0] |
---|
305 | |
---|
306 | if nodeslist.count( host ) == 0: |
---|
307 | |
---|
308 | for translate_pattern in BATCH_HOST_TRANSLATE: |
---|
309 | |
---|
310 | translate_orig = translate_pattern.split( '/' )[1] |
---|
311 | translate_new = translate_pattern.split( '/' )[2] |
---|
312 | |
---|
313 | host = re.sub( translate_orig, translate_new, host ) |
---|
314 | |
---|
315 | if not host in nodeslist: |
---|
316 | |
---|
317 | nodeslist.append( host ) |
---|
318 | |
---|
319 | if DETECT_TIME_DIFFS: |
---|
320 | |
---|
321 | # If a job start if later than our current date, |
---|
322 | # that must mean the Torque server's time is later |
---|
323 | # than our local time. |
---|
324 | |
---|
325 | if int(start_timestamp) > int( int(self.cur_time) + int(self.timeoffset) ): |
---|
326 | |
---|
327 | self.timeoffset = int( int(start_timestamp) - int(self.cur_time) ) |
---|
328 | |
---|
329 | elif status == 'Q': |
---|
330 | start_timestamp = '' |
---|
331 | count_mynodes = 0 |
---|
332 | numeric_node = 1 |
---|
333 | |
---|
334 | for node in mynoderequest.split( '+' ): |
---|
335 | |
---|
336 | nodepart = node.split( ':' )[0] |
---|
337 | |
---|
338 | for letter in nodepart: |
---|
339 | |
---|
340 | if letter not in string.digits: |
---|
341 | |
---|
342 | numeric_node = 0 |
---|
343 | |
---|
344 | if not numeric_node: |
---|
345 | count_mynodes = count_mynodes + 1 |
---|
346 | else: |
---|
347 | count_mynodes = count_mynodes + int( nodepart ) |
---|
348 | |
---|
349 | nodeslist = count_mynodes |
---|
350 | else: |
---|
351 | start_timestamp = '' |
---|
352 | nodeslist = '' |
---|
353 | |
---|
354 | myAttrs = { } |
---|
355 | myAttrs['name'] = name |
---|
356 | myAttrs['queue'] = queue |
---|
357 | myAttrs['owner'] = owner |
---|
358 | myAttrs['requested_time'] = requested_time |
---|
359 | myAttrs['requested_memory'] = requested_memory |
---|
360 | myAttrs['ppn'] = ppn |
---|
361 | myAttrs['status'] = status |
---|
362 | myAttrs['start_timestamp'] = start_timestamp |
---|
363 | myAttrs['reported'] = str( int( int( self.cur_time ) + int( self.timeoffset ) ) ) |
---|
364 | myAttrs['nodes'] = nodeslist |
---|
365 | myAttrs['domain'] = string.join( socket.getfqdn().split( '.' )[1:], '.' ) |
---|
366 | myAttrs['poll_interval'] = TORQUE_POLL_INTERVAL |
---|
367 | |
---|
368 | if self.jobDataChanged( jobs, job_id, myAttrs ) and myAttrs['status'] in [ 'R', 'Q' ]: |
---|
369 | jobs[ job_id ] = myAttrs |
---|
370 | |
---|
371 | #debug_msg( 10, printTime() + ' job %s state changed' %(job_id) ) |
---|
372 | |
---|
373 | for id, attrs in jobs.items(): |
---|
374 | |
---|
375 | if id not in jobs_processed: |
---|
376 | |
---|
377 | # This one isn't there anymore; toedeledoki! |
---|
378 | # |
---|
379 | del jobs[ id ] |
---|
380 | |
---|
381 | return jobs |
---|
382 | |
---|
383 | def submitJobData( self, jobs ): |
---|
384 | """Submit job info list""" |
---|
385 | |
---|
386 | self.dp.multicastGmetric( 'MONARCH-HEARTBEAT', str( int( int( self.cur_time ) + int( self.timeoffset ) ) ) ) |
---|
387 | |
---|
388 | # Now let's spread the knowledge |
---|
389 | # |
---|
390 | for jobid, jobattrs in jobs.items(): |
---|
391 | |
---|
392 | gmetric_val = self.compileGmetricVal( jobid, jobattrs ) |
---|
393 | |
---|
394 | for val in gmetric_val: |
---|
395 | self.dp.multicastGmetric( 'MONARCH-JOB-' + jobid, val ) |
---|
396 | |
---|
397 | def makeNodeString( self, nodelist ): |
---|
398 | """Make one big string of all hosts""" |
---|
399 | |
---|
400 | node_str = None |
---|
401 | |
---|
402 | for node in nodelist: |
---|
403 | if not node_str: |
---|
404 | node_str = node |
---|
405 | else: |
---|
406 | node_str = node_str + ';' + node |
---|
407 | |
---|
408 | return node_str |
---|
409 | |
---|
410 | def compileGmetricVal( self, jobid, jobattrs ): |
---|
411 | """Create a val string for gmetric of jobinfo""" |
---|
412 | |
---|
413 | appendList = [ ] |
---|
414 | appendList.append( 'name=' + jobattrs['name'] ) |
---|
415 | appendList.append( 'queue=' + jobattrs['queue'] ) |
---|
416 | appendList.append( 'owner=' + jobattrs['owner'] ) |
---|
417 | appendList.append( 'requested_time=' + jobattrs['requested_time'] ) |
---|
418 | |
---|
419 | if jobattrs['requested_memory'] != '': |
---|
420 | appendList.append( 'requested_memory=' + jobattrs['requested_memory'] ) |
---|
421 | |
---|
422 | if jobattrs['ppn'] != '': |
---|
423 | appendList.append( 'ppn=' + jobattrs['ppn'] ) |
---|
424 | |
---|
425 | appendList.append( 'status=' + jobattrs['status'] ) |
---|
426 | |
---|
427 | if jobattrs['start_timestamp'] != '': |
---|
428 | appendList.append( 'start_timestamp=' + jobattrs['start_timestamp'] ) |
---|
429 | |
---|
430 | appendList.append( 'reported=' + jobattrs['reported'] ) |
---|
431 | appendList.append( 'poll_interval=' + str( jobattrs['poll_interval'] ) ) |
---|
432 | appendList.append( 'domain=' + jobattrs['domain'] ) |
---|
433 | |
---|
434 | if jobattrs['status'] == 'R': |
---|
435 | if len( jobattrs['nodes'] ) > 0: |
---|
436 | appendList.append( 'nodes=' + self.makeNodeString( jobattrs['nodes'] ) ) |
---|
437 | elif jobattrs['status'] == 'Q': |
---|
438 | appendList.append( 'nodes=' + str(jobattrs['nodes']) ) |
---|
439 | |
---|
440 | return self.makeAppendLists( appendList ) |
---|
441 | |
---|
442 | def makeAppendLists( self, append_list ): |
---|
443 | """ |
---|
444 | Divide all values from append_list over strings with a maximum |
---|
445 | size of 1400 |
---|
446 | """ |
---|
447 | |
---|
448 | app_lists = [ ] |
---|
449 | |
---|
450 | mystr = None |
---|
451 | |
---|
452 | for val in append_list: |
---|
453 | |
---|
454 | if not mystr: |
---|
455 | mystr = val |
---|
456 | else: |
---|
457 | if not self.checkValAppendMaxSize( mystr, val ): |
---|
458 | mystr = mystr + ' ' + val |
---|
459 | else: |
---|
460 | # Too big, new appenlist |
---|
461 | app_lists.append( mystr ) |
---|
462 | mystr = val |
---|
463 | |
---|
464 | app_lists.append( mystr ) |
---|
465 | |
---|
466 | return app_lists |
---|
467 | |
---|
468 | def checkValAppendMaxSize( self, val, text ): |
---|
469 | """Check if val + text size is not above 1400 (max msg size)""" |
---|
470 | |
---|
471 | # Max frame size of a udp datagram is 1500 bytes |
---|
472 | # removing misc header and gmetric stuff leaves about 1300 bytes |
---|
473 | # |
---|
474 | if len( val + text ) > 900: |
---|
475 | return 1 |
---|
476 | else: |
---|
477 | return 0 |
---|
478 | |
---|
479 | def printJobs( self, jobs ): |
---|
480 | """Print a jobinfo overview""" |
---|
481 | |
---|
482 | for name, attrs in self.jobs.items(): |
---|
483 | |
---|
484 | print 'job %s' %(name) |
---|
485 | |
---|
486 | for name, val in attrs.items(): |
---|
487 | |
---|
488 | print '\t%s = %s' %( name, val ) |
---|
489 | |
---|
490 | def printJob( self, jobs, job_id ): |
---|
491 | """Print job with job_id from jobs""" |
---|
492 | |
---|
493 | print 'job %s' %(job_id) |
---|
494 | |
---|
495 | for name, val in jobs[ job_id ].items(): |
---|
496 | |
---|
497 | print '\t%s = %s' %( name, val ) |
---|
498 | |
---|
499 | def daemon( self ): |
---|
500 | """Run as daemon forever""" |
---|
501 | |
---|
502 | # Fork the first child |
---|
503 | # |
---|
504 | pid = os.fork() |
---|
505 | if pid > 0: |
---|
506 | sys.exit(0) # end parent |
---|
507 | |
---|
508 | # creates a session and sets the process group ID |
---|
509 | # |
---|
510 | os.setsid() |
---|
511 | |
---|
512 | # Fork the second child |
---|
513 | # |
---|
514 | pid = os.fork() |
---|
515 | if pid > 0: |
---|
516 | sys.exit(0) # end parent |
---|
517 | |
---|
518 | # Go to the root directory and set the umask |
---|
519 | # |
---|
520 | os.chdir('/') |
---|
521 | os.umask(0) |
---|
522 | |
---|
523 | sys.stdin.close() |
---|
524 | sys.stdout.close() |
---|
525 | sys.stderr.close() |
---|
526 | |
---|
527 | os.open('/dev/null', 0) |
---|
528 | os.dup(0) |
---|
529 | os.dup(0) |
---|
530 | |
---|
531 | self.run() |
---|
532 | |
---|
533 | def run( self ): |
---|
534 | """Main thread""" |
---|
535 | |
---|
536 | while ( 1 ): |
---|
537 | |
---|
538 | self.jobs = self.getJobData( self.jobs ) |
---|
539 | self.submitJobData( self.jobs ) |
---|
540 | time.sleep( TORQUE_POLL_INTERVAL ) |
---|
541 | |
---|
542 | def printTime( ): |
---|
543 | """Print current time/date in human readable format for log/debug""" |
---|
544 | |
---|
545 | return time.strftime("%a, %d %b %Y %H:%M:%S") |
---|
546 | |
---|
547 | def debug_msg( level, msg ): |
---|
548 | """Print msg if at or above current debug level""" |
---|
549 | |
---|
550 | if (DEBUG_LEVEL >= level): |
---|
551 | sys.stderr.write( msg + '\n' ) |
---|
552 | |
---|
553 | def main(): |
---|
554 | """Application start""" |
---|
555 | |
---|
556 | if not processArgs( sys.argv[1:] ): |
---|
557 | sys.exit( 1 ) |
---|
558 | |
---|
559 | gather = DataGatherer() |
---|
560 | if DAEMONIZE: |
---|
561 | gather.daemon() |
---|
562 | else: |
---|
563 | gather.run() |
---|
564 | |
---|
565 | # w00t someone started me |
---|
566 | # |
---|
567 | if __name__ == '__main__': |
---|
568 | main() |
---|