| 1 |
#!/usr/bin/python2 |
| 2 |
# lcg-info-dynamic-scheduler |
| 3 |
# $Id: lcg-info-dynamic-scheduler.cin,v 4.4 2007/08/14 12:05:58 templon Exp $ |
| 4 |
# J. A. Templon, NIKHEF/PDP 2005 |
| 5 |
|
| 6 |
# plugin for gip framework. generates Glue values associated |
| 7 |
# with the scheduler, like FreeSlots or *ResponseTime |
| 8 |
|
| 9 |
# first set up logging |
| 10 |
|
| 11 |
import sys, logging |
| 12 |
import syslog |
| 13 |
|
| 14 |
# note: this class works around the bug in the python logging |
| 15 |
# package fixed in patch #642974. If the version of python-logging |
| 16 |
# is upgraded, this class can be replaced by logging.SysLogHandler |
| 17 |
|
| 18 |
class SLHandler(logging.Handler): |
| 19 |
def __init__(self, ident, logopt=0, facility=syslog.LOG_USER): |
| 20 |
logging.Handler.__init__(self) |
| 21 |
self.ident = ident |
| 22 |
self.logopt = logopt |
| 23 |
self.facility = facility |
| 24 |
self.mappings = { |
| 25 |
logging.DEBUG: syslog.LOG_DEBUG, |
| 26 |
logging.INFO: syslog.LOG_INFO, |
| 27 |
logging.WARN: syslog.LOG_WARNING, |
| 28 |
logging.ERROR: syslog.LOG_ERR, |
| 29 |
logging.CRITICAL: syslog.LOG_CRIT, |
| 30 |
} |
| 31 |
|
| 32 |
def encodeLevel(self, level): |
| 33 |
return self.mappings.get(level, syslog.LOG_INFO) |
| 34 |
|
| 35 |
def emit(self, record): |
| 36 |
syslog.openlog(self.ident, self.logopt, self.facility) |
| 37 |
msg = self.format(record) |
| 38 |
prio = self.encodeLevel(record.levelno) |
| 39 |
syslog.syslog(prio, msg) |
| 40 |
syslog.closelog() |
| 41 |
|
| 42 |
logging.getLogger("").setLevel(logging.INFO) |
| 43 |
# syslog handler |
| 44 |
shdlr = SLHandler("lcg-info-dynamic-scheduler") |
| 45 |
logging.getLogger("").addHandler(shdlr) |
| 46 |
# stderr handler |
| 47 |
stdrhdlr = logging.StreamHandler() |
| 48 |
fmt=logging.Formatter("%(asctime)s lcg-info-dynamic-scheduler:" |
| 49 |
+ " %(message)s","%F %T") |
| 50 |
logging.getLogger("").addHandler(stdrhdlr) |
| 51 |
stdrhdlr.setFormatter(fmt) |
| 52 |
|
| 53 |
def usage(): |
| 54 |
print "Usage: lcg-info-dynamic-scheduler -c <cfg_file>" |
| 55 |
|
| 56 |
import sys |
| 57 |
|
| 58 |
def abort_without_output(msg): |
| 59 |
logging.error(msg) |
| 60 |
logging.error("Exiting without output, GIP will use static values") |
| 61 |
sys.exit(2) |
| 62 |
|
| 63 |
import getopt |
| 64 |
import string |
| 65 |
|
| 66 |
try: |
| 67 |
opts, args = getopt.getopt(sys.argv[1:], "c:", |
| 68 |
["config="]) |
| 69 |
except getopt.GetoptError: |
| 70 |
# print help information and exit: |
| 71 |
emsg = "Error parsing command line: " + string.join(sys.argv) |
| 72 |
usage() |
| 73 |
abort_without_output(emsg) |
| 74 |
|
| 75 |
cfgfile = None |
| 76 |
|
| 77 |
for o, a in opts: |
| 78 |
if o in ("-c", "--config"): |
| 79 |
cfgfile = a |
| 80 |
|
| 81 |
if not cfgfile: |
| 82 |
usage() |
| 83 |
abort_without_output("No config file specified.") |
| 84 |
|
| 85 |
try: |
| 86 |
fp = open(cfgfile) |
| 87 |
except IOError: |
| 88 |
abort_without_output("Error opening config file " + cfgfile) |
| 89 |
|
| 90 |
import commands |
| 91 |
import ConfigParser |
| 92 |
config = ConfigParser.ConfigParser() |
| 93 |
config.readfp(fp) |
| 94 |
|
| 95 |
# note: vomap is only intended for the cases where voname != groupname |
| 96 |
# hence at the bottom, we throw away all entries for which voname = groupname |
| 97 |
# this is the "default scenario" in the code, adding such entries to the |
| 98 |
# vomap "exception list" causes problems later on. |
| 99 |
|
| 100 |
if config.has_option('Main','vomap'): |
| 101 |
s = config.get('Main','vomap') |
| 102 |
vomap = dict() |
| 103 |
lines = s.split('\n') |
| 104 |
for l in lines: |
| 105 |
if len(l) > 0: |
| 106 |
t = l.split(':') |
| 107 |
group = t[0].strip() |
| 108 |
vo = t[1].strip() |
| 109 |
if vo != group : vomap[group] = vo |
| 110 |
|
| 111 |
else: |
| 112 |
vomap = dict() # empty dict, later 'if grp in vomap.keys()' => if False |
| 113 |
|
| 114 |
if config.has_option('Scheduler','vo_max_jobs_cmd'): |
| 115 |
cmd = config.get('Scheduler','vo_max_jobs_cmd') |
| 116 |
(stat,out) = commands.getstatusoutput(cmd) |
| 117 |
if stat: |
| 118 |
abort_without_output( |
| 119 |
"VO max jobs backend command returned nonzero exit status") |
| 120 |
try: |
| 121 |
pcaps = eval(out, {"__builtins__" : {}}) # namespace at closes security hole |
| 122 |
except SyntaxError: |
| 123 |
abort_without_output('vo max slot command output in wrong format') |
| 124 |
else: |
| 125 |
pcaps = dict() # empty dict, will trigger use of defaults later. |
| 126 |
|
| 127 |
# apply group -> vo mappings to process cap dict |
| 128 |
# sgroup (source group) is original from input files, |
| 129 |
# targvo (target vo) is what we want to replace it with |
| 130 |
|
| 131 |
#DEBUG opcaps = pcaps.copy() |
| 132 |
|
| 133 |
for sgroup in vomap.keys(): |
| 134 |
targ = vomap[sgroup] |
| 135 |
if sgroup in pcaps.keys(): |
| 136 |
if targ in pcaps.keys(): |
| 137 |
pcaps[targ] = pcaps[targ] + pcaps[sgroup] |
| 138 |
else: |
| 139 |
pcaps[targ] = pcaps[sgroup] |
| 140 |
del pcaps[sgroup] |
| 141 |
|
| 142 |
#DEBUGkl = opcaps.keys() + pcaps.keys() |
| 143 |
#DEBUGkl.sort() |
| 144 |
#DEBUGukl = [] |
| 145 |
#DEBUGfor k in kl: |
| 146 |
#DEBUG if k not in ukl: ukl.append(k) |
| 147 |
#DEBUGfor k in ukl: |
| 148 |
#DEBUG v1 = -1 |
| 149 |
#DEBUG v2 = -1 |
| 150 |
#DEBUG if k in opcaps.keys(): |
| 151 |
#DEBUG v1 = opcaps[k] |
| 152 |
#DEBUG if k in pcaps.keys(): |
| 153 |
#DEBUG v2 = pcaps[k] |
| 154 |
#DEBUG |
| 155 |
#DEBUG print "%10s %4d %4d" % (k, v1, v2) |
| 156 |
|
| 157 |
if not config.has_option('LRMS','lrms_backend_cmd'): |
| 158 |
abort_without_output("Spec for lrms backend cmd missing in config file") |
| 159 |
|
| 160 |
# read cfg file to find spec for static ldif file |
| 161 |
|
| 162 |
if not config.has_option('Main','static_ldif_file'): |
| 163 |
abort_without_output("Spec for static ldif file missing in config file") |
| 164 |
|
| 165 |
lines = open(config.get('Main','static_ldif_file'),'r').readlines() |
| 166 |
|
| 167 |
# get the dns from the ldif file along with their VOs. |
| 168 |
# first do the VOView dns |
| 169 |
|
| 170 |
voviewlist = [ ] |
| 171 |
qname = None |
| 172 |
import re |
| 173 |
jmpatt = re.compile(r'^GlueCEUniqueID=[-a-zA-Z0-9.:]+/[a-z]+-[a-z]+-([a-z_A-Z0-9]+)$') |
| 174 |
|
| 175 |
class CEView(object): |
| 176 |
def __init__(self, dn): |
| 177 |
self.dn = dn |
| 178 |
# if this is not present, then it's a GlueCE object |
| 179 |
# if it is present, it's a GlueVOView object |
| 180 |
self.localID = None |
| 181 |
self.queuename = None |
| 182 |
self.ACBRs = list() |
| 183 |
def getCEUniqueID(self): |
| 184 |
if self.localID == None: |
| 185 |
return self.dn |
| 186 |
else: |
| 187 |
loc = self.dn.find(",") + 1 |
| 188 |
return self.dn[loc:] |
| 189 |
CEUniqueID = property(getCEUniqueID, |
| 190 |
doc="UniqueID of CE to which this view belongs") |
| 191 |
|
| 192 |
dndict = dict() # key: dn string (not including leading "dn: ") |
| 193 |
thisview = None |
| 194 |
|
| 195 |
for line in lines: |
| 196 |
|
| 197 |
s = line.strip() |
| 198 |
if s.find('dn:') == 0 and s.find('GlueCEUniqueID') > 0: |
| 199 |
dn = s[3:].strip() |
| 200 |
thisview = CEView(dn) |
| 201 |
if s.find('GlueVOViewLocalID') >= 0: |
| 202 |
loc1 = s.find('=') + 1 |
| 203 |
loc2 = s.find(',GlueCEUniqueID') |
| 204 |
id = s[loc1:loc2].strip() |
| 205 |
thisview.localID = id |
| 206 |
elif s.find('GlueCEName') == 0: |
| 207 |
qname = s.split(':')[1].strip() |
| 208 |
thisview.queuename = qname |
| 209 |
elif s.find('GlueCEAccessControlBaseRule:') == 0: |
| 210 |
loc = s.find(':') + 1 |
| 211 |
rule = s[loc:].split(":") |
| 212 |
if rule[0].strip() in ['VO', 'VOMS']: |
| 213 |
thisview.ACBRs.append( ( rule[0].strip(), rule[1].strip() ) ) |
| 214 |
|
| 215 |
elif s == '' and thisview : |
| 216 |
dndict[thisview.dn] = thisview |
| 217 |
thisview = None # reset, about to enter new block. |
| 218 |
|
| 219 |
# now find the queues that go with the VOViews |
| 220 |
for d in dndict.keys(): |
| 221 |
view = dndict[d] |
| 222 |
if view.localID: # then it's a VOView |
| 223 |
view.queuename = dndict[view.CEUniqueID].queuename |
| 224 |
|
| 225 |
DEBUG = 0 |
| 226 |
|
| 227 |
if DEBUG: |
| 228 |
print "dumping parse tree for static ldif file" |
| 229 |
for d in dndict.keys(): |
| 230 |
print "For dn:", d |
| 231 |
print " LocalID: ", dndict[d].localID |
| 232 |
print " CEUniqueID: ", dndict[d].CEUniqueID |
| 233 |
print " Queue Name:", dndict[d].queuename |
| 234 |
print " ACBRs: ", dndict[d].ACBRs |
| 235 |
|
| 236 |
def tconv(timeval): # convert time to desired units |
| 237 |
timeval = min(timeval, 2146060842) |
| 238 |
return repr( int(timeval) ) # seconds->printable |
| 239 |
|
| 240 |
cmd = config.get('LRMS','lrms_backend_cmd') |
| 241 |
|
| 242 |
(stat,out) = commands.getstatusoutput(cmd) |
| 243 |
if stat: |
| 244 |
abort_without_output("LRMS backend command returned nonzero exit status") |
| 245 |
|
| 246 |
lrmlines = out.split('\n') |
| 247 |
|
| 248 |
sys.path.append('@MODDIR@') |
| 249 |
import lrms |
| 250 |
bq = lrms.Server() |
| 251 |
|
| 252 |
# generate a cache of indices returned from possible lrms queries |
| 253 |
# this makes a big speed diff when there are many (thousands) of jobs |
| 254 |
# first we need to find all combinations VO/fqan, queue, state |
| 255 |
|
| 256 |
ql = [] |
| 257 |
vol = [] |
| 258 |
|
| 259 |
for dn in dndict.keys(): |
| 260 |
view = dndict[dn] |
| 261 |
if not view.localID : continue |
| 262 |
if view.queuename not in ql : ql.append(view.queuename) |
| 263 |
if view.ACBRs[-1][1] not in vol : vol.append(view.ACBRs[-1][1]) |
| 264 |
|
| 265 |
sl = ['running', 'queued'] |
| 266 |
|
| 267 |
# now construct an empty cache with all the index keys, ready to fill. |
| 268 |
|
| 269 |
scache = { } |
| 270 |
|
| 271 |
for v in vol: |
| 272 |
for s in sl: |
| 273 |
indstr = lrms.filt2str({'group':v, 'state': s}) |
| 274 |
scache[indstr] = {} |
| 275 |
|
| 276 |
for q in ql: |
| 277 |
for s in sl: |
| 278 |
indstr = lrms.filt2str({'queue': q, 'state': s}) |
| 279 |
scache[indstr] = {} |
| 280 |
for v in vol: |
| 281 |
indstr = lrms.filt2str({'queue': q, 'state': s, 'group': v}) |
| 282 |
scache[indstr] = {} |
| 283 |
|
| 284 |
#DEBUG print scache.keys() |
| 285 |
|
| 286 |
# traverse the job list, filling the index cache as we go. |
| 287 |
|
| 288 |
for line in lrmlines: |
| 289 |
s = line.strip() |
| 290 |
f = s.split() |
| 291 |
if s[0] == '{' and s[-1] == '}': # looks like a dict |
| 292 |
nj = lrms.Job(eval(s, {"__builtins__" : {}})) |
| 293 |
grp = nj.get('group') |
| 294 |
if grp in vomap.keys(): |
| 295 |
nj.set('group',vomap[grp]) |
| 296 |
jid = nj.get('jobid') |
| 297 |
bq.addjob(jid,nj) |
| 298 |
tst = nj.get('state') |
| 299 |
if tst not in sl: continue |
| 300 |
tgr = nj.get('group') |
| 301 |
tq = nj.get('queue') |
| 302 |
if tgr in vol and tq in ql: |
| 303 |
dlis = [ |
| 304 |
{'group': tgr, 'state': tst}, |
| 305 |
{'queue': tq, 'state': tst}, |
| 306 |
{'queue': tq, 'state': tst, 'group': tgr} |
| 307 |
] |
| 308 |
else: |
| 309 |
if tgr in vol: |
| 310 |
dlis = [ |
| 311 |
{'group': tgr, 'state': tst}, |
| 312 |
] |
| 313 |
elif tq in ql: |
| 314 |
dlis = [ |
| 315 |
{'queue': tq, 'state': tst}, |
| 316 |
{'queue': tq, 'state': tst, 'group': tgr} |
| 317 |
] |
| 318 |
else: |
| 319 |
continue |
| 320 |
|
| 321 |
for td in dlis : |
| 322 |
indstr = lrms.filt2str(td) |
| 323 |
if indstr not in scache.keys(): |
| 324 |
if DEBUG : print "huh??", indstr |
| 325 |
scache[indstr] = {} |
| 326 |
scache[lrms.filt2str(td)][jid] = nj |
| 327 |
elif len(f) == 2: |
| 328 |
if f[0] == 'nactive' : bq.slotsUp = int(f[1]) |
| 329 |
elif f[0] == 'nfree' : bq.slotsFree = int(f[1]) |
| 330 |
elif f[0] == 'now' : bq.now = int(f[1]) |
| 331 |
elif f[0] == 'schedCycle' : bq.schedCycle = int(f[1]) |
| 332 |
else: |
| 333 |
logging.warn("invalid input in result of lrms plugin command " + cmd) |
| 334 |
logging.warn("invalid input: " + s) |
| 335 |
else: |
| 336 |
logging.warn("invalid input in result of lrms plugin command " + cmd) |
| 337 |
logging.warn("invalid input: " + s) |
| 338 |
|
| 339 |
bq.__scache__ = scache |
| 340 |
|
| 341 |
# for k in scache.keys(): |
| 342 |
# print k, len(scache[k]) |
| 343 |
|
| 344 |
if config.has_option('Scheduler','cycle_time'): |
| 345 |
bq.schedCycle = config.getint('Scheduler','cycle_time') |
| 346 |
|
| 347 |
# sys.exit(0) |
| 348 |
|
| 349 |
from EstTT import * |
| 350 |
wq = TTEstimator(WaitTime()) |
| 351 |
|
| 352 |
# first we do the VOViews |
| 353 |
# sort the keys : needed to make sure output is reproducible for test harness :( |
| 354 |
|
| 355 |
dnk = dndict.keys() |
| 356 |
dnk.sort() |
| 357 |
|
| 358 |
for dn in dnk: |
| 359 |
|
| 360 |
view = dndict[dn] |
| 361 |
if not view.localID: continue # we only want VOViews here |
| 362 |
q = view.queuename |
| 363 |
if len(view.ACBRs) > 1: |
| 364 |
logging.warn("For dn: " + dn) |
| 365 |
logging.warn("detected multiple ACBRs:") |
| 366 |
for vo in view.ACBRs: |
| 367 |
logging.warn(repr(vo)) |
| 368 |
logging.warn("only using the last one!") |
| 369 |
vo = view.ACBRs[-1][1] |
| 370 |
|
| 371 |
if DEBUG: print "'" + vo + "'" |
| 372 |
|
| 373 |
print "dn: " + dn |
| 374 |
print "GlueVOViewLocalID: " + view.localID |
| 375 |
if DEBUG: |
| 376 |
for acbr in view.ACBRs: |
| 377 |
print "GlueCEAccessControlBaseRule: " + acbr[0] + ":" + acbr[1] |
| 378 |
|
| 379 |
# this is tricky: reporting job counts in queues cares about queues, |
| 380 |
# but scheduling doesn't as far as I know. So we need four job |
| 381 |
# counts, running and waiting, both in the current queue and for all |
| 382 |
# queues. |
| 383 |
|
| 384 |
# first find what is happening in this queue (q) to report the |
| 385 |
# WaitingJobs RunningJobs and TotalJobs |
| 386 |
|
| 387 |
|
| 388 |
qrund = {'state': 'running', 'group': vo, 'queue': q} |
| 389 |
qwaitd = {'state': 'queued', 'group': vo, 'queue': q} |
| 390 |
|
| 391 |
runningJobs = bq.matchingJobs(qrund) ; nrun = len(runningJobs) |
| 392 |
waitingJobs = bq.matchingJobs(qwaitd); nwait = len(waitingJobs) |
| 393 |
if nwait > 0: |
| 394 |
qwt = waitingJobs[0].get('maxwalltime') |
| 395 |
else: |
| 396 |
qwt = -1 |
| 397 |
|
| 398 |
print "GlueCEStateRunningJobs: " + repr(nrun) |
| 399 |
print "GlueCEStateWaitingJobs: " + repr(nwait) |
| 400 |
print "GlueCEStateTotalJobs: " + repr(nrun + nwait) |
| 401 |
|
| 402 |
# now we find the same numbers, but integrated across all queues |
| 403 |
# since the scheduler applies process caps without regard to queues. |
| 404 |
|
| 405 |
rund = {'state': 'running', 'group': vo} |
| 406 |
waitd = {'state': 'queued', 'group': vo} |
| 407 |
|
| 408 |
runningJobs = bq.matchingJobs(rund) ; nrun = len(runningJobs) |
| 409 |
waitingJobs = bq.matchingJobs(waitd); nwait = len(waitingJobs) |
| 410 |
|
| 411 |
if nwait > 0: |
| 412 |
freeslots = 0 |
| 413 |
else: |
| 414 |
if vo in pcaps.keys(): |
| 415 |
headroom = max(pcaps[vo] - nrun,0) |
| 416 |
freeslots = min(bq.slotsFree,headroom) |
| 417 |
else: |
| 418 |
freeslots = bq.slotsFree |
| 419 |
|
| 420 |
ert = wq.estimate(bq,vo,debug=0) |
| 421 |
print "GlueCEStateFreeJobSlots: " + repr(freeslots) |
| 422 |
print "GlueCEStateEstimatedResponseTime: " + tconv(ert) |
| 423 |
if nwait == 0: |
| 424 |
wrt = 2 * ert |
| 425 |
else: |
| 426 |
if qwt < 0 : |
| 427 |
qwt = waitingJobs[0].get('maxwalltime') |
| 428 |
wrt = qwt * nwait |
| 429 |
|
| 430 |
print "GlueCEStateWorstResponseTime: " + \ |
| 431 |
tconv((max(wrt,bq.schedCycle))) |
| 432 |
print |
| 433 |
|
| 434 |
# now do CEs (no VOs) ... this should be done 'right' some day |
| 435 |
# and merged with the code above. Want to get it out quick |
| 436 |
# (sigh, i am mortal) so ... |
| 437 |
|
| 438 |
for dn in dnk: |
| 439 |
|
| 440 |
view = dndict[dn] |
| 441 |
if view.localID: continue # only want CEs, they have no localID |
| 442 |
q = view.queuename |
| 443 |
vo = '' # vo-independent |
| 444 |
|
| 445 |
print "dn: " + dn |
| 446 |
|
| 447 |
# for VO-independent numbers we only care about what is in the current |
| 448 |
# queue. so we only need one set of stats |
| 449 |
|
| 450 |
# first find what is happening in this queue (q) to report the |
| 451 |
# WaitingJobs RunningJobs and TotalJobs |
| 452 |
|
| 453 |
qrund = {'state': 'running', 'queue': q} |
| 454 |
qwaitd = {'state': 'queued', 'queue': q} |
| 455 |
|
| 456 |
runningJobs = bq.matchingJobs(qrund) ; nrun = len(runningJobs) |
| 457 |
waitingJobs = bq.matchingJobs(qwaitd); nwait = len(waitingJobs) |
| 458 |
|
| 459 |
if nwait > 0: |
| 460 |
freeslots = 0 |
| 461 |
else: |
| 462 |
freeslots = bq.slotsFree |
| 463 |
|
| 464 |
ert = wq.estimate(bq,vo,debug=0) |
| 465 |
print "GlueCEStateFreeJobSlots: " + repr(freeslots) |
| 466 |
print "GlueCEStateFreeCPUs: " + repr(nrun + nwait) # backw compat. |
| 467 |
print "GlueCEStateRunningJobs: " + repr(nrun) |
| 468 |
print "GlueCEStateWaitingJobs: " + repr(nwait) |
| 469 |
print "GlueCEStateTotalJobs: " + repr(nrun + nwait) |
| 470 |
|
| 471 |
wq.strategy.queue = q |
| 472 |
ert = wq.estimate(bq,vo,debug=0) |
| 473 |
print "GlueCEStateEstimatedResponseTime: " + tconv(ert) |
| 474 |
if nwait == 0: |
| 475 |
wrt = 2 * ert |
| 476 |
else: |
| 477 |
wrt = waitingJobs[0].get('maxwalltime') * nwait |
| 478 |
|
| 479 |
print "GlueCEStateWorstResponseTime: " + \ |
| 480 |
tconv((max(wrt,bq.schedCycle))) |
| 481 |
if DEBUG: print 'DEBUG: q', q |
| 482 |
print |
| 483 |
|
| 484 |
### Local Variables: *** |
| 485 |
### mode: python *** |
| 486 |
### End: *** |
| 487 |
|
Parent Directory
| 
Revision Log