class MPISubmission:
def __init__(self, np, execfname ):
self.debug = DEBUG
self.np = np
self.mpdPort = ''
self.nfsTmp = NFS_PREFIX + gethostname(True)
self.execfname = execfname
self.servthread = None
self.rand = genRandom()
self.tmpPath = self._create_tmpdir()
self.submitFile = TemplateFile( '' , SUBM_FNAME, self.tmpPath,
SUBM_FNAME + '_' + self.rand )
self.clientFile = TemplateFile( '' , CLNT_FNAME, self.tmpPath,
CLNT_FNAME + '_' + self.rand )
self.hostfname = self.tmpPath + 'hosts_' + self.rand
self.keyfname = self.tmpPath + self.rand + '.key'
self.hostlist = [] # store a list of [username, ip, port]
self.env = os.environ
self.env['PATH'] = self.nfsTmp + '/' + MPDBIN_PATH + ':' + self.env['PATH']
# remove tmp dir and all its content
def _rm_tmpdir(self, dst='.'):
tpath = dst + '/tmp' + self.rand
try:
if os.access( tpath, os.F_OK) == True:
shutil.rmtree( tpath )
except os.error as e:
sys.exit('Error: cannot remove temp directory - ' + e )
def _create_tmpdir(self, dst='.'):
try:
os.makedirs( dst + '/tmp' + self.rand )
except os.error as e:
sys.exit('Error: cannot create temp directory - ' + e )
return 'tmp' + self.rand + '/'
def _start_mpd(self):
# make sure no other mpds are running
subprocess.call( 'mpdallexit', env=self.env, stdout=FNULL, stderr=FNULL)
time.sleep(1.0) # wait for old mpd ring to be torn down
# starting server's mpd
mpdconf_path = os.getcwd()+'/'+self.tmpPath[:-1]
createMpdConf( self.rand, mpdconf_path )
self.env['MPD_CONF_FILE'] = mpdconf_path + '/.mpd.conf'
subprocess.Popen(['mpd', '--daemon'], env=self.env)
time.sleep(1.0) # wait for mpd to fully start
# determine mpd listening port from mpdtrace output
process = subprocess.Popen(['mpdtrace', '-l'], stdout=subprocess.PIPE, env=self.env)
process.wait()
traceout = process.communicate()[0]
port = extractPort(traceout)
if not port.isdigit():
sys.exit('Error starting mpd : ' + traceout )
self.mpdPort = port
def start(self):
# Copy exe file into local ganfs directory
self._create_tmpdir( self.nfsTmp )
shutil.copy2( self.execfname, self.nfsTmp + '/' + self.tmpPath )
self._start_mpd() # start local mpd
self._prepare_submission_files() # prepare client script and condor submit file
self._gen_ssh_keys() # generate ssh key pairs
# start a listening server
self.servthread = Thread(target=CallbackServ(self.np - 1, self.hostfname, PORT).serv)
self.servthread.setDaemon(True)
self.servthread.start()
if self.debug:
print "submit condor with file " + str(self.submitFile)
p = subprocess.call(['condor_submit', str(self.submitFile) ] )
if p != 0:
sys.exit('Error: condor_submit return ' + str(p))
# if the submission is successful, wait for the server
print 'Waiting for ' + str(self.np-1) + ' workers to response ....',
sys.stdout.flush()
self.servthread.join()
print 'finished'
self._read_hosts_info() # read info from the collected hosts
# Waiting for mpd ring to be ready
print 'Waiting for mpd ring to be ready .......',
limit = 120
retry = 0
while retry < limit :
time.sleep(1.0) # wait
# testing mpd connection
process = subprocess.Popen(['mpdtrace', '-l'], env=self.env,
stdout=subprocess.PIPE )
process.wait()
trace = process.communicate()[0]
retry += 1
port = extractPort(trace)
num = len( trace.split('\n') )
if port.isdigit() and (num == self.np + 1):
print 'done'
if self.debug:
print '\nMPD trace:\n' + trace
break
# Check whether mpdtrace return enough mpd nodes
if len(trace.split('\n')) < self.np + 1 :
print 'failed'
subprocess.call( 'mpdallexit', env=self.env, stdout=FNULL, stderr=FNULL)
subprocess.call(['condor_rm', '-all'])
sys.exit('Error: not enough mpd nodes in the ring')
# Run mpi job
execdir = self.nfsTmp + '/' + self.tmpPath
subprocess.call(['mpiexec', '-n', str(self.np),
execdir + self.execfname.split('/')[-1]], env=self.env)
# mpi job is finished
for host in self.hostlist: # notify all workers
hostserv = xmlrpclib.Server( "http://" + host[0] + ":" + host[4] )
hostserv.terminate()
subprocess.call( 'mpdallexit', env=self.env,
stdout=FNULL, stderr=FNULL) # tear down mpd ring
self._rm_tmpdir( self.nfsTmp ) # remove exec file from ganfs dir
self._rm_tmpdir() # remove temp directory
def _prepare_submission_files(self):
# Prepare condor submission file
self.submitFile.prepare_file( [ ['<q.np>', str(self.np-1)],
['<ssh.pub.key>', self.tmpPath + AUTH_FNAME ],
['<fullpath.client.script>', str(self.clientFile) ],
['<output.dir>', OUTPUT_DIR + '/' ],
['<client.script>', self.clientFile.out_fname() ] ] )
# Prepare client side python script
if self.debug:
print 'serv ipop ip = ' + get_ipop_ip()
self.clientFile.prepare_file(
[ [ '<serv.ip>', get_ipop_ip() ],
[ '<serv.port>', str(PORT) ],
[ '<mpd.port>', self.mpdPort ],
[ '<mpd.path>', self.nfsTmp +'/'+ MPDBIN_PATH],
[ '<rand>', self.rand ] ], True )
def _read_hosts_info(self):
with open( self.hostfname, 'r') as hostf:
for line in hostf:
self.hostlist.append( line.rstrip().split(":") )
hostf.close()
if self.debug:
print '\nhost list:'
for host in self.hostlist:
print host
def _create_mpd_hosts(self):
with open( self.tmpPath + 'mpd.hosts', 'w') as mpdhostf:
for host in self.hostlist:
# write hostname and number of cpus only
mpdhostf.write( host[0] + ':' + host[1] + '\n' )
mpdhostf.close()
# Generate ssh key pairs for MPI ring
def _gen_ssh_keys(self):
argstr = str.split( "ssh-keygen -q -t rsa" )
argstr.extend( ["-N", ''] )
argstr.extend( ["-f", self.keyfname] )
argstr.extend( ["-C", self.rand] )
p = subprocess.call( argstr )
if p != 0:
sys.exit('Error: ssh-keygen return ' + str(p))
# copy the public key file into 'authorized_keys' file for client's sshd
shutil.copyfile( self.keyfname + '.pub', self.tmpPath + AUTH_FNAME )
class HadoopCluster:
def __init__(self, np ):
self.debug = DEBUG
self.np = np
self.nfsTmp = NFS_PREFIX + gethostname(True)
self.servthread = None
self.rand = 'hadoopAAA'
self.tmpPath = self._create_tmpdir()
self.submitFile = TemplateFile( '' , SUBM_FNAME, self.tmpPath,
SUBM_FNAME + '_' + self.rand )
self.clientFile = TemplateFile( '' , CLNT_FNAME, self.tmpPath,
CLNT_FNAME + '_' + self.rand )
self.hostfname = self.tmpPath + 'hosts_' + self.rand
self.hostlist = [] # store a list of [username, ip, port]
self.env = os.environ
self.env['HADOOP_HEAPSIZE'] = str(128)
self.fullTmpPath = (os.getcwd() + '/' + self.tmpPath)[:-1]
self.env['HADOOP_CONF_DIR'] = self.fullTmpPath
self.env['HADOOP_PID_DIR'] = self.fullTmpPath
self.env['HADOOP_LOG_DIR'] = self.fullTmpPath
self.env['HADOOP_HOME'] = self.nfsTmp + '/' + HADP_PATH
self.env['JAVA_HOME'] = self.nfsTmp + '/' + JAVA_PATH
# testing and add hadoop/bin into PATH
binPath = self.nfsTmp + '/' + HADP_PATH + '/bin'
if string.find( self.env['PATH'], binPath ) == -1:
self.env['PATH'] = binPath + ':' + self.env['PATH']
# remove tmp dir and all its content
def _rm_tmpdir(self, dst='.'):
tpath = dst + '/tmp' + self.rand
if os.path.isdir( tpath ):
try:
if os.access( tpath, os.F_OK) == True:
shutil.rmtree( tpath )
except os.error as e:
sys.exit('Error: cannot remove temp directory - ' + e )
def _create_tmpdir(self, dst='.'):
tmpdir = dst + '/tmp' + self.rand
if not os.path.isdir( tmpdir ):
try:
os.makedirs( tmpdir )
except os.error as e:
sys.exit('Error: cannot create temp directory - ' + e )
return 'tmp' + self.rand + '/'
def _stop_hadoop(self):
subprocess.call( ['hadoop-daemon.sh', 'stop', 'datanode'],
env=self.env, stdout=FNULL, stderr=FNULL)
subprocess.call( ['hadoop-daemon.sh', 'stop', 'namenode'],
env=self.env, stdout=FNULL, stderr=FNULL)
subprocess.call( ['hadoop-daemon.sh', 'stop', 'tasktracker'],
env=self.env, stdout=FNULL, stderr=FNULL)
subprocess.call( ['hadoop-daemon.sh', 'stop', 'jobtracker'],
env=self.env, stdout=FNULL, stderr=FNULL)
def _start_hadoop_dfs(self):
# starting server's namenode
p1 = subprocess.Popen( ['echo', 'Y'], stdout=subprocess.PIPE, stderr=FNULL)
subprocess.call( ['hdfs', 'namenode', '-format'], env=self.env,
stdin=p1.stdout, stdout=FNULL, stderr=FNULL )
p1.stdout.close()
print 'Starting a namenode (this may take a while) .... ',
sys.stdout.flush()
subprocess.Popen( ['hadoop-daemon.sh', 'start', 'namenode'],
env=self.env, stdout=FNULL, stderr=FNULL)
# use -report to wait for namenode to finish starting up
subprocess.call(['hdfs', 'dfsadmin' ,'-report'],
env=self.env, stdout=FNULL, stderr=FNULL )
print 'done'
# Start local datanode
print 'Starting a local datanode'
subprocess.Popen( ['hadoop-daemon.sh', 'start', 'datanode'],
env=self.env, stdout=FNULL, stderr=FNULL)
def _start_hadoop_mapred(self):
# Start local jobtracker
print 'Starting a local jobtracker'
subprocess.Popen( ['hadoop-daemon.sh', 'start', 'jobtracker'],
env=self.env, stdout=FNULL, stderr=FNULL)
# Start local tasktracker
print 'Starting a local tasktracker\n'
subprocess.Popen( ['hadoop-daemon.sh', 'start', 'tasktracker'],
env=self.env, stdout=FNULL, stderr=FNULL)
def start(self):
# Create conf file into local temp directory
confFile = TemplateFile( '' , CONF_TMP_FNAME, self.tmpPath, CONF_FNAME )
confFile.prepare_file([ ['<namenode.hostname>', gethostname()] ])
hdfsFile = TemplateFile( '', HDFS_TMP_FNAME, self.tmpPath, HDFS_FNAME )
hdfsFile.prepare_file([ ['<name.dir>', self.fullTmpPath + '/name' ],
['<data.dir>', self.fullTmpPath + '/data' ] ] )
mprdFile = TemplateFile( '', MPRD_TMP_FNAME, self.tmpPath, MPRD_FNAME )
mprdFile.prepare_file([ ['<jobtracker.hostname>', gethostname()],
['<jobtracker.port>', str(MPRD_PORT) ] ])
# Prepare and copy hadoop-env.sh into local temp directory
envFile = TemplateFile('', HENV_TMP_FNAME, self.tmpPath, HENV_FNAME )
envFile.prepare_file([ ['<java.home>', self.env['JAVA_HOME']],
['<hadoop.home>', self.env['HADOOP_HOME']],
['<conf.dir>', self.env['HADOOP_CONF_DIR']],
['<pid.dir>', self.env['HADOOP_PID_DIR']],
['<log.dir>', self.env['HADOOP_LOG_DIR']],
['<path>', self.env['PATH']] ], True )
self._start_hadoop_dfs() # start local hadoop namenode & datanode
self._prepare_submission_files() # prepare client script and condor submit file
# start a listening server
self.servthread = Thread(target=CallbackServ(self.np-1, self.hostfname, PORT).serv)
self.servthread.setDaemon(True)
self.servthread.start()
if self.debug:
print "submit condor with file " + str(self.submitFile)
p = subprocess.call(['condor_submit', str(self.submitFile) ] )
if p != 0:
sys.exit('Error: condor_submit return ' + str(p))
# if the submission is successful, wait for the server
print 'Waiting for ' + str(self.np-1) + ' workers to response ....',
sys.stdout.flush()
self.servthread.join()
print 'finished'
self._read_hosts_info() # read info from the collected hosts
# Waiting for hadoop cluster to be ready
print '\nWaiting for ' + str(self.np) + ' datanodes to be ready ',
print '(This may take a while) .... ',
sys.stdout.flush()
limit = 180
retry = 0
while retry < limit :
time.sleep(1.0)
# testing hadoop cluster
process = subprocess.Popen(['hdfs', 'dfsadmin' ,'-report'],
env=self.env, stdout=subprocess.PIPE, stderr=FNULL )
process.wait()
trace = process.communicate()[0]
retry += 1
# extract datanodes count
start = trace.find("Datanodes available:")
end = trace.find("(", start )
count = '0'
if start > 0 and end > 0:
count = trace[start:end].split()[-1]
if count.isdigit() and (int(count) == self.np ):
print 'success'
break
# Check whether a report return enough datanodes
if retry >= limit :
print 'fail'
self.stop()
sys.exit('Timeout: not enough datanodes in the cluster')
self._start_hadoop_mapred()
print 'Attention: Please source ' + self.tmpPath + HENV_FNAME +' before using hadoop.\n'
def stop(self):
self._read_hosts_info() # read info from tmp dir
if len(self.hostlist) == 0 :
sys.exit('Error: no existing hadoop cluster info' )
for host in self.hostlist: # notify all workers
hostserv = xmlrpclib.Server( "http://" + host[0] + ":" + host[3] )
try:
hostserv.terminate()
except socket.error:
print 'host ' + host[0] + ':' + host[3] + ' is not respending.'
self._stop_hadoop() # stop hadoop
self._rm_tmpdir() # remove temp directory
def _prepare_submission_files(self):
# Prepare condor submission file
self.submitFile.prepare_file( [ ['<q.np>', str(self.np-1)],
['<fullpath.client.script>', str(self.clientFile) ],
['<output.dir>', OUTPUT_DIR + '/' ],
['<core.config.file>', self.tmpPath + CONF_FNAME ],
['<mprd.config.file>', self.tmpPath + MPRD_FNAME ],
['<hdfs.config.tmp.file>', HDFS_TMP_FNAME ],
['<env.config.file>', self.tmpPath + HENV_FNAME ],
['<client.script>', self.clientFile.out_fname() ] ] )
# Prepare client side python script
if self.debug:
print 'serv ipop ip = ' + get_ipop_ip()
self.clientFile.prepare_file(
[ [ '<serv.ip>', get_ipop_ip() ],
[ '<serv.port>', str(PORT) ],
[ '<hadp.path>', self.nfsTmp + '/' + HADP_PATH ],
[ '<java.path>', self.nfsTmp + '/' + JAVA_PATH ],
[ '<hdfs.config.file>', HDFS_FNAME ],
[ '<hdfs.config.tmp.file>', HDFS_TMP_FNAME ],
[ '<rand>', self.rand ] ], True )
def _read_hosts_info(self):
self.hostlist = []
with open( self.hostfname, 'r') as hostf:
for line in hostf:
self.hostlist.append( line.rstrip().split(":") )
hostf.close()
if self.debug:
print '\nhost list:'
for host in self.hostlist:
print host