| 1 |
#!/usr/bin/env python |
| 2 |
from torque_utils import pbsnodes |
| 3 |
|
| 4 |
# settings below appropriate for knal nodes (32 cores) |
| 5 |
# ... want to keep num of draining nodes |
| 6 |
# to MAXFREE / 7 since giving back a node could be expensive. |
| 7 |
# it might already be running several mc jobs |
| 8 |
|
| 9 |
MAXDRAIN = 7 # max num of nodes allowed to drain |
| 10 |
MAXFREE = 49 # max num of free slots to tolerate |
| 11 |
CANDIDATE_NODES = [ 'wn-knal-0%02d.farm.nikhef.nl' % (n) for n in range(1,19) ] |
| 12 |
MCQUEUE = 'atlasmc' |
| 13 |
|
| 14 |
# settings below more appropriate for smrt nodes (8 core) |
| 15 |
# here giving a node back doesn't really matter, it wasn't |
| 16 |
# running any mc jobs yet anyway. better to grab lots |
| 17 |
# of nodes and give back the ones that drain the slowest if |
| 18 |
# too many slots are free. |
| 19 |
|
| 20 |
# MAXDRAIN = 16 # max num of nodes allowed to drain |
| 21 |
# MAXFREE = 49 # max num of free slots to tolerate |
| 22 |
# CANDIDATE_NODES = [ 'wn-smrt-0%02d.farm.nikhef.nl' % (n) for n in range(1,73) ] |
| 23 |
|
| 24 |
UNDERPOP_NODES_FILE = ".nodes_with_too_few_jobs" |
| 25 |
TORQUE = "stro.nikhef.nl" |
| 26 |
|
| 27 |
import os |
| 28 |
import pickle |
| 29 |
import sys |
| 30 |
import subprocess |
| 31 |
|
| 32 |
def getmcjobinfo(nodes): |
| 33 |
|
| 34 |
# input is the list of nodes for which to get info |
| 35 |
# function returns tuple (jobs_waiting, node_info) |
| 36 |
# jobs_waiting boolean value : are there MC jobs waiting to be run (state Q)? (independent of nodes!) |
| 37 |
# node_info: list of tuples (wn_name, wn, num_running_mc, num_free_slots) |
| 38 |
# where wn is the node object for the machine with name wn_name, num_running_mc is the number of |
| 39 |
# mc jobs running on that node, and num_free_slots is the number of unused job slots on that node |
| 40 |
|
| 41 |
usedkeys = ['egroup', 'jobid', 'queue', 'job_state', 'euser', 'exec_host' ] |
| 42 |
|
| 43 |
import torqueJobs |
| 44 |
import torqueAttMappers as tam |
| 45 |
|
| 46 |
import tempfile |
| 47 |
qsfname = tempfile.mktemp(".txt","qsmf",os.environ['HOME']+"/tmp") |
| 48 |
|
| 49 |
import time |
| 50 |
os.system('/usr/bin/qstat -f @' + TORQUE + ' > ' + qsfname) |
| 51 |
now = time.mktime(time.localtime()) |
| 52 |
|
| 53 |
jlist = torqueJobs.qs_parsefile(qsfname) |
| 54 |
|
| 55 |
os.system('mv ' + qsfname + ' ' + os.environ['HOME'] + '/tmp/qstat.last.txt') |
| 56 |
|
| 57 |
def mapatts(indict,inkeys): |
| 58 |
sdict = tam.sub_dict(indict,inkeys) |
| 59 |
odict = dict() |
| 60 |
for k in sdict.keys(): |
| 61 |
if k in tam.tfl and sdict[k]: |
| 62 |
secs = tam.hms(sdict[k]) |
| 63 |
sdict[k] = secs/3600. |
| 64 |
if k in tam.mfl and sdict[k]: |
| 65 |
mebi = tam.memconvert(sdict[k]) |
| 66 |
sdict[k] = mebi |
| 67 |
if k in tam.tfl2 and sdict[k]: |
| 68 |
secs = tam.tconv(sdict[k]) |
| 69 |
sdict[k] = secs |
| 70 |
elif k == 'job_state': |
| 71 |
statelett = sdict[k] |
| 72 |
if statelett in ['Q','W']: |
| 73 |
sdict[k] = 'queued' |
| 74 |
elif statelett in ['R','E']: |
| 75 |
sdict[k] = 'running' |
| 76 |
elif k == 'exec_host' and sdict[k]: |
| 77 |
termpos = sdict[k].find('/') |
| 78 |
wnstring = sdict[k][:termpos] |
| 79 |
sdict[k] = wnstring |
| 80 |
if sdict[k]: |
| 81 |
odict[k] = sdict[k] |
| 82 |
return odict |
| 83 |
|
| 84 |
mcjoblist = list() |
| 85 |
waitingJobs = False |
| 86 |
for j in jlist: |
| 87 |
if j['queue'] == MCQUEUE: |
| 88 |
newj = mapatts(j,usedkeys) |
| 89 |
mcjoblist.append(newj) |
| 90 |
if newj['job_state'] == 'queued': |
| 91 |
waitingJobs = True |
| 92 |
|
| 93 |
# find out how many running mc jobs each node has |
| 94 |
|
| 95 |
running_mc = dict() |
| 96 |
for n in nodes: |
| 97 |
running_mc[n.name] = 0 |
| 98 |
for j in mcjoblist: |
| 99 |
if j['job_state'] =='running': |
| 100 |
wn = j['exec_host'] |
| 101 |
if wn in running_mc.keys(): |
| 102 |
running_mc[wn] += 1 |
| 103 |
rawlist = list() |
| 104 |
for n in nodes: |
| 105 |
rawlist.append( (n.name, n, running_mc[n.name], n.freeCpu) ) |
| 106 |
from operator import itemgetter |
| 107 |
slist = sorted(rawlist, key=lambda x: (x[2], x[3])) |
| 108 |
|
| 109 |
return waitingJobs, slist |
| 110 |
|
| 111 |
def remove_from_mc_pool(node): |
| 112 |
# print "adding el6 tag back to node", node.name |
| 113 |
if "el6" not in node.properties : |
| 114 |
proc = subprocess.Popen(['/usr/bin/qmgr'], |
| 115 |
stdin=subprocess.PIPE, |
| 116 |
stdout=subprocess.PIPE,) |
| 117 |
proc.stdin.write( 'set node ' + node.name + \ |
| 118 |
' properties += el6\n' ) |
| 119 |
proc.stdin.write( 'print node ' + node.name + \ |
| 120 |
' properties\n' ) |
| 121 |
out = proc.communicate()[0] |
| 122 |
# print out |
| 123 |
if "mc" in node.properties : |
| 124 |
proc = subprocess.Popen(['/usr/bin/qmgr'], |
| 125 |
stdin=subprocess.PIPE, |
| 126 |
stdout=subprocess.PIPE,) |
| 127 |
proc.stdin.write( 'set node ' + node.name + \ |
| 128 |
' properties -= mc\n' ) |
| 129 |
proc.stdin.write( 'print node ' + node.name + \ |
| 130 |
' properties\n' ) |
| 131 |
out = proc.communicate()[0] |
| 132 |
# print out |
| 133 |
|
| 134 |
def add_to_mc_pool(node): |
| 135 |
# print "node props b4:", grabbed_node.properties |
| 136 |
if "el6" in node.properties : |
| 137 |
proc = subprocess.Popen(['/usr/bin/qmgr'], |
| 138 |
stdin=subprocess.PIPE, |
| 139 |
stdout=subprocess.PIPE,) |
| 140 |
proc.stdin.write( 'set node ' + node.name + \ |
| 141 |
' properties -= el6\n' ) |
| 142 |
proc.stdin.write( 'print node ' + node.name + \ |
| 143 |
' properties\n' ) |
| 144 |
out = proc.communicate()[0] |
| 145 |
|
| 146 |
if "mc" not in node.properties : |
| 147 |
proc = subprocess.Popen(['/usr/bin/qmgr'], |
| 148 |
stdin=subprocess.PIPE, |
| 149 |
stdout=subprocess.PIPE,) |
| 150 |
proc.stdin.write( 'set node ' + node.name + \ |
| 151 |
' properties += mc\n' ) |
| 152 |
proc.stdin.write( 'print node ' + node.name + \ |
| 153 |
' properties\n' ) |
| 154 |
out = proc.communicate()[0] |
| 155 |
|
| 156 |
import optparse |
| 157 |
|
| 158 |
usage = "usage: %prog [-d debug_level] [-n]" |
| 159 |
|
| 160 |
p = optparse.OptionParser(description="Monitor state of multicore pool and adjust size as needed", |
| 161 |
usage=usage) |
| 162 |
|
| 163 |
p.add_option("-n",action="store_true",dest="noopt",default=False, |
| 164 |
help="don't do anything, just print what would have been done") |
| 165 |
p.add_option("-l",action="store",dest="logfile",default=None, |
| 166 |
help="log actions and information to LOGFILE (default stdout)") |
| 167 |
p.add_option("-L",action="store",dest="loglevel",default="INFO", |
| 168 |
help="print messages of LOGLEVEL (DEBUG, INFO, WARNING, ..., default INFO") |
| 169 |
|
| 170 |
import logging |
| 171 |
|
| 172 |
opts, args = p.parse_args() |
| 173 |
|
| 174 |
# assuming loglevel is bound to the string value obtained from the |
| 175 |
# command line argument. Convert to upper case to allow the user to |
| 176 |
# specify --log=DEBUG or --log=debug |
| 177 |
|
| 178 |
numeric_level = getattr(logging, opts.loglevel.upper(), None) |
| 179 |
if not isinstance(numeric_level, int): |
| 180 |
raise ValueError('Invalid log level: %s' % loglevel) |
| 181 |
if opts.logfile: |
| 182 |
logging.basicConfig(format='%(asctime)s %(levelname)s %(message)s', |
| 183 |
level=numeric_level,filename=opts.logfile) |
| 184 |
else: |
| 185 |
logging.basicConfig(format='%(asctime)s %(levelname)s %(message)s', |
| 186 |
level=numeric_level) |
| 187 |
|
| 188 |
if os.path.isfile(UNDERPOP_NODES_FILE): |
| 189 |
pkl_file = open(UNDERPOP_NODES_FILE,'rb') |
| 190 |
nodes_too_few_jobs_last_run = pickle.load(pkl_file) |
| 191 |
pkl_file.close() |
| 192 |
else: |
| 193 |
nodes_too_few_jobs_last_run = list() |
| 194 |
|
| 195 |
logging.debug("nodes from pickle file, marked from last run: " + \ |
| 196 |
repr(nodes_too_few_jobs_last_run) ) |
| 197 |
|
| 198 |
wnodes = pbsnodes() |
| 199 |
mcnodes = list() |
| 200 |
waiting = None |
| 201 |
node_tuples = None |
| 202 |
|
| 203 |
for node in wnodes: |
| 204 |
if node.name in CANDIDATE_NODES and node.state.count('offline') == 0: |
| 205 |
mcnodes.append(node) |
| 206 |
|
| 207 |
draining_slots = 0 |
| 208 |
mcdedicated = list() |
| 209 |
draining_nodes = 0 |
| 210 |
|
| 211 |
for node in mcnodes: |
| 212 |
if 'el6' not in node.properties and 'mc' in node.properties : |
| 213 |
mcdedicated.append(node) |
| 214 |
draining_slots += node.freeCpu |
| 215 |
if node.freeCpu > 0 : draining_nodes += 1 |
| 216 |
|
| 217 |
# check for dedicated nodes with too few jobs |
| 218 |
|
| 219 |
nodes_with_too_few_jobs = list() |
| 220 |
|
| 221 |
for node in mcdedicated: |
| 222 |
logging.debug(node.name + " has " + repr(node.freeCpu) + " free slots") |
| 223 |
if node.freeCpu > 7: |
| 224 |
nodes_with_too_few_jobs.append(node) |
| 225 |
|
| 226 |
logging.debug("there are " + repr(len(nodes_with_too_few_jobs)) + \ |
| 227 |
" nodes with too few jobs") |
| 228 |
|
| 229 |
nodes_consistently_underpopulated = list() |
| 230 |
for n in nodes_with_too_few_jobs: |
| 231 |
if n.name in nodes_too_few_jobs_last_run: |
| 232 |
nodes_consistently_underpopulated.append(n) |
| 233 |
|
| 234 |
undcount = len(nodes_consistently_underpopulated) |
| 235 |
if undcount > 0: |
| 236 |
logging.debug("there are " + repr(undcount) + \ |
| 237 |
" nodes with too few jobs that were also marked last run" ) |
| 238 |
|
| 239 |
import random |
| 240 |
|
| 241 |
removed_a_node = False |
| 242 |
|
| 243 |
if undcount > 0: |
| 244 |
logging.info("nodes consistently underpopulated:") |
| 245 |
for n in nodes_consistently_underpopulated: |
| 246 |
logging.info(" " + n.name) |
| 247 |
if undcount > 1: |
| 248 |
remcount = undcount / 2 # number to remove |
| 249 |
logging.info("going to remove " + repr(remcount) + " from mc pool") |
| 250 |
else: |
| 251 |
remcount = 1 |
| 252 |
|
| 253 |
# find out how many running mc jobs each node has |
| 254 |
|
| 255 |
waiting, node_tuples = getmcjobinfo(nodes_consistently_underpopulated) |
| 256 |
|
| 257 |
for node_t in node_tuples[:remcount]: |
| 258 |
nname, nnode, running_mc, unused_slots = node_t |
| 259 |
logging.info("dumped %d empty slots, %d mc jobs on %s" % \ |
| 260 |
(unused_slots, running_mc, nname) ) |
| 261 |
if not opts.noopt : remove_from_mc_pool(nnode) |
| 262 |
nodes_with_too_few_jobs.remove(nnode) |
| 263 |
nodes_consistently_underpopulated.remove(nnode) |
| 264 |
removed_a_node = True |
| 265 |
|
| 266 |
namelist = list() |
| 267 |
if len(nodes_with_too_few_jobs) > 0: |
| 268 |
logging.debug("placing %d nodes " % (len(nodes_with_too_few_jobs)) + \ |
| 269 |
"with too few jobs on list for next run:") |
| 270 |
|
| 271 |
for n in nodes_with_too_few_jobs: |
| 272 |
logging.debug(n.name) |
| 273 |
namelist.append(n.name) |
| 274 |
if not opts.noopt: |
| 275 |
pkl_file = open(UNDERPOP_NODES_FILE,'wb') |
| 276 |
pickle.dump(namelist,pkl_file) |
| 277 |
pkl_file.close() |
| 278 |
if len(nodes_with_too_few_jobs) > 0 or removed_a_node : |
| 279 |
sys.exit(0) |
| 280 |
|
| 281 |
# if we survive to this point, all nodes 'dedicated' are being efficiently |
| 282 |
# used. are we draining less than the configured max? |
| 283 |
|
| 284 |
logging.debug("There are " + repr(len(mcdedicated)) + " dedicated nodes and " + \ |
| 285 |
repr(draining_slots) + " unused slots") |
| 286 |
if (MAXFREE - draining_slots) >= 7: |
| 287 |
logging.debug("%d unused slots are permitted, so %d more" % (MAXFREE, MAXFREE - draining_slots)) |
| 288 |
logging.debug("headroom of 7+ slots means we can try to grab another node") |
| 289 |
if draining_nodes < MAXDRAIN : |
| 290 |
# first check if there are actually any waiting jobs to run; if not makes no sense to grab a node. |
| 291 |
waiting, node_tuples = getmcjobinfo(mcdedicated) |
| 292 |
if not waiting: |
| 293 |
logging.debug("No waiting jobs found, nothing to do") |
| 294 |
sys.exit(0) |
| 295 |
logging.debug("max %d node(s) with unused slots permitted, %d seen" % \ |
| 296 |
(MAXDRAIN, draining_nodes)) |
| 297 |
logging.debug("there are also waiting jobs: try to grab another node") |
| 298 |
# build a list of candidates to grab |
| 299 |
candidate_nodes = list() |
| 300 |
for n in mcnodes: |
| 301 |
if n not in mcdedicated: |
| 302 |
candidate_nodes.append(n) |
| 303 |
logging.debug("found %d candidate nodes to dedicate to mc" % len(candidate_nodes)) |
| 304 |
if len(candidate_nodes) < 1: |
| 305 |
logging.debug("no more nodes, bailing out, nothing more I can do") |
| 306 |
sys.exit(0) |
| 307 |
grabbed_node=random.choice(candidate_nodes) |
| 308 |
logging.info("%s added to mc node pool" % (grabbed_node.name)) |
| 309 |
if not opts.noopt: add_to_mc_pool(grabbed_node) |
| 310 |
else: |
| 311 |
logging.debug("There are %d nodes with unused slots (draining)" % (draining_nodes)) |
| 312 |
logging.debug("This equals or exceeds the configured max of %d" % (MAXDRAIN)) |
| 313 |
logging.debug("Doing nothing now") |
| 314 |
|
| 315 |
elif draining_slots > MAXFREE: |
| 316 |
|
| 317 |
# find out how many running mc jobs each node has |
| 318 |
|
| 319 |
waiting, node_tuples = getmcjobinfo(mcdedicated) |
| 320 |
|
| 321 |
slots_to_recover = draining_slots - MAXFREE |
| 322 |
logging.info("unused slot limit (%d) exceeded by %d " \ |
| 323 |
% (MAXFREE, slots_to_recover) + ": remove node(s) from mc pool") |
| 324 |
slots_recovered = 0 |
| 325 |
while slots_recovered < slots_to_recover: |
| 326 |
node_t = node_tuples.pop(0) |
| 327 |
nname, nnode, running_mc, unused_slots = node_t |
| 328 |
if unused_slots > 0: |
| 329 |
logging.info("dumped %d empty slots, %d mc jobs on %s" % \ |
| 330 |
(unused_slots, running_mc, nname) ) |
| 331 |
if not opts.noopt : remove_from_mc_pool(nnode) |
| 332 |
slots_recovered += unused_slots |
| 333 |
else: |
| 334 |
logging.debug("%d unused slots of allowed %d" % (draining_slots, MAXFREE)) |
| 335 |
logging.debug("difference is %d which is less than 7" % \ |
| 336 |
(MAXFREE - draining_slots) ) |
| 337 |
logging.debug("so: doing nothing now.") |
Parent Directory
| 
Revision Log