def load(self):
pastHosts = Session.query(Host).all()
print ("INFRASTRUCTURE-LOAD: LOADING INFORMATION FROM PAST EXECUTIONS")
print (" Desired LRMS: " + self.lrms)
#obtain present hosts prom gridway
#PILOTS
pastPilots = Session.query(Pilot).all()
print (" Deleting all past pilots from database")
for pilot in pastPilots:
base.Session.delete(pilot)
presentHosts = []
GWHosts = obtainGWResources()
if GWHosts == []:
print("Error when parsing host information file, employing information from past executions")
self.hosts = pastHosts
return
#load XML to memory and extract data from hosts
for hostInfo in GWHosts:
hostName = hostInfo.getElementsByTagName("HOSTNAME")[0].firstChild.data #TODO: remove "unicode" from TEXT
#for every found host, check if it existed on a previous execution
for host in pastHosts:
if host.hostname.strip().lower() == hostName.strip().lower():
#in the case of pilot jobs, only employ the ones with FREENODECOUNT > 0
if host.lrms =="jobmanager-pilot":
try:
freeNodeCount = int(hostInfo.getElementsByTagName("FREENODECOUNT")[0].firstChild.data)
except:
freeNodeCount = 0
if not freeNodeCount > 0:
continue;
presentHosts.append(host)
break # Si no pones este break, y está repetido en memoria, se cargan todos
#now, if only hosts with a certain LRMS are desired, we remove the rest
if self.lrms != None:
for host in presentHosts:
if host.lrms != self.lrms:
print (" Removing host " + host.hostname + ", wrong LRMS found: " + host.lrms)
presentHosts.remove(host)
else:
print (" Keeping host " + host.hostname)
#the hosts that we will first employ are the resulting ones
self.hosts = presentHosts
def updateStatus (self, gridTasks):
print ("UPDATE STATUS")
GWHosts = obtainGWResources()
if GWHosts ==[]:
print("Error when parsing host information file, employing information from past executions")
return GWHosts
#PILOTS
pastPilots = Session.query(Pilot).all()
presentPilots = []
#load XML to memory and extract data from hosts
for resource in GWHosts:
hostname = resource.getElementsByTagName("HOSTNAME")[0].firstChild.data.strip().lower() #TODO: remove "unicode" from TEXT
try:
foundLrms = resource.getElementsByTagName("LRMS_NAME")[0].firstChild.data.strip().lower()
except:
print ("Could not obtain resource LRMS, skipping it")
continue
if foundLrms != "jobmanager-pilot":
continue
#ahora son todos pilots
try:
freeNodeCount = int(resource.getElementsByTagName("FREENODECOUNT")[0].firstChild.data)
except:
freeNodeCount = 0
if not freeNodeCount > 0:
continue;
for pilot in pastPilots:
if pilot.hostname == hostname:
presentPilots.append(pilot)
break
#the pilots that we will first employ are the resulting ones
self.pilots = presentPilots
def createInfrastructureTasks(self, infrastructureTasks):
print ("---------------------")
print ("---------------------")
print ("---------------------")
print ("CREATE INFRASTRUCTURE TASKS")
hostsToProfile = []
hostList = obtainGWResources()
for hostInfo in hostList:
hostName = hostInfo.getElementsByTagName("HOSTNAME")[0].firstChild.data #TODO: remove "unicode" from TEXT
try:
foundArch = hostInfo.getElementsByTagName("ARCH")[0].firstChild.data
except:
foundArch=""
try:
foundCpuMHz = int(hostInfo.getElementsByTagName("CPU_MHZ")[0].firstChild.data)
except:
foundCpuMHz = 0
try:
foundLrms = hostInfo.getElementsByTagName("LRMS_NAME")[0].firstChild.data
except:
foundLrms = None
try:
freeNodeCount = int(hostInfo.getElementsByTagName("FREENODECOUNT")[0].firstChild.data)
except:
freeNodeCount = 0
if foundLrms != None:
if foundLrms == "jobmanager-pilot":
#solo tenemos en cuenta los pilots con al menos un slot disponible
if not freeNodeCount > 0:
continue
#if a certain LRMS is desired, remove the hosts with a different one
if self.lrms != None:
if foundLrms != self.lrms:
continue
#if host is unknown, create a profiling task
currentHost = self.getHost(hostName)
if currentHost == None:
newHost = Host(hostName, arch=foundArch, cpuMHz = foundCpuMHz, lrms=foundLrms)
self.hosts.append(newHost)
hostsToProfile.append(newHost)
#store new host on databae (faiulre resistance
Session.add(newHost)
#if information has changed, update host information
elif (currentHost.arch != foundArch) or (currentHost.cpuMHz != foundCpuMHz):
#TODO: pensar que hacer aqui. habria que eliminar el viejo o solo sobreescribir la información? Si se elimina el viejo, que pasa con las tareas ahí ejecutadas? No es trivial
currentHost.arch = foundArch
currentHost.cpuMHz = foundCpuMHz
if currentHost.lrms == None:
currentHost.lrms = foundLrms
hostsToProfile.append(currentHost)
Session.add(currentHost)
elif currentHost.shouldBeProfiled():
if currentHost.lrms == None:
currentHost.lrms = foundLrms
hostsToProfile.append(currentHost)
#print("Host profiling: submission of 1 tasks per host")
hostProfilingTasks = [ExecutionManager.createHostProfilingTask(host)
for host in hostsToProfile
for i in range(1)]
siteTasks = []
for task in hostProfilingTasks:
found=False
for gridTask in infrastructureTasks:
if gridTask.host.hostname == task.host.hostname:
found=True
break
if not found:
siteTasks.append(task)
#Esto es para el primer experimento de montera + gwpilot
#queremos tener pilots funcionando, así que los arranco con esto
if self.lrms=="jobmanager-pilot":
print ("creating fake profiling tasks")
existingFakeTasks = len([task for task in infrastructureTasks if task.host.hostname=="" and task.status != "PENDING"])
existingGoodPilots = len (self.getGoodHosts())
existingProfilingTasks = len(hostProfilingTasks)
#fakeTasksToCreate = base.maxRunningTasks - (existingFakeTasks + existingGoodPilots + existingProfilingTasks)
fakeTasksToCreate = base.maxRunningTasks - existingFakeTasks
print (" Desired tasks: " + str(base.maxRunningTasks))
print (" Existing fake tasks: " + str(existingFakeTasks))
print (" Existing good pilots: " + str(existingGoodPilots))
print (" created: " + str(fakeTasksToCreate))
emptyHost = FakeHost()
fakeHostProfilingTasks = [ExecutionManager.createWakeUpask(emptyHost)
for i in range(fakeTasksToCreate)]
siteTasks+=fakeHostProfilingTasks
return siteTasks
def getGoodHosts(self):
print ("GET GOOD HOSTS.")
pastHostsList = [host for host in self.hosts if host.successfulExecutions > 0 ]
hostList = obtainGWResources()
#load XML to memory, and extract data from hosts
presentHostNames = []
for hostInfo in hostList:
hostName = hostInfo.getElementsByTagName("HOSTNAME")[0].firstChild.data #TODO: remove "unicode" from TEXT
presentHostNames.append(hostName)
for host in pastHostsList:
found = False
for name in presentHostNames:
if name == host.hostname:
found = True
break
if not found:
pastHostsList.remove(host)
if host.lrms != self.lrms:
pastHostsList.remove(host)
#banning failure hosts
for host in pastHostsList:
bannedTime = None
if host.failedProfilings < 0:
continue
elif host.failedProfilings == 0 and host.successfulExecutions > 0:
continue
#24 primeros fallos: banning de una hora por fallo
if host.failedProfilings < 24:
bannedTime =timedelta(hours=host.failedProfilings)
elif host.failedProfilings >= 24:
bannedTime = timedelta(days=7)
if bannedTime != None:
if (datetime.now() - bannedTime) < host.lastFailedProfiling:
pastHostsList.remove(host)
# print (" Host "+ host.hostname + " is banned due to failures for " + str(bannedTime))
#ban hosts with no whetstones, what means no profiling
hostsWithWhetstones = []
for host in pastHostsList:
if host.getWhetstones() > 1:
hostsWithWhetstones.append(host)
# else:
# print (" Host "+ host.hostname + " is banned, no whetstone info")
#Un poco chapuza esto, pero bueno
defList=[]
if self.lrms == "jobmanager-pilot":
print (" filtering hosts.")
for host in hostsWithWhetstones:
if host.lrms == self.lrms:
defList.append(host)
print (" keeped host " + host.hostname + " with lrms=" + host.lrms)
else:
print (" deleted host " + host.hostname + " with lrms=" + host.lrms)
else:
defList = hostsWithWhetstones
print (" Returning list of size " + str(len(defList)))
for host in defList:
print (" " + host.hostname)
return defList
def createInfrastructureTasks(self, infrastructureTasks):
print ("-------------------")
print ("-------------------")
print ("createInfrastructureTasks- NewPilotInfrastructure")
# self.showHosts()
hostList = obtainGWResources()
hostsToProfile = []
print ("Analyzing resources ")
for hostInfo in hostList:
hostName = hostInfo.getElementsByTagName("HOSTNAME")[0].firstChild.data.strip().lower() #TODO: remove "unicode" from TEXT
whetstones=0
try:
foundArch = hostInfo.getElementsByTagName("ARCH")[0].firstChild.data
except:
foundArch=""
try:
foundCpuMHz = int(hostInfo.getElementsByTagName("CPU_MHZ")[0].firstChild.data)
except:
foundCpuMHz = 0
try:
foundLrms = hostInfo.getElementsByTagName("LRMS_NAME")[0].firstChild.data.strip().lower()
except:
foundLrms = None
print ("Could not find LRMS for host " + hostName + ", skipping it")
continue
try:
freeNodeCount = int(hostInfo.getElementsByTagName("FREENODECOUNT")[0].firstChild.data)
except:
freeNodeCount = 0
if foundLrms == "jobmanager-pilot":
#solo tenemos en cuenta los pilots con al menos un slot disponible
if not freeNodeCount > 0:
continue
username = os.getenv("USER")
genericStringArgs = hostInfo.getElementsByTagName("GENERIC_VAR_STR")
for node in genericStringArgs:
if node.attributes['NAME'].value =="PILOT_REAL_HOSTNAME":
workerNode = node.attributes['VALUE'].value.strip().lower()
if node.attributes['NAME'].value =="PILOT_REAL_RESOURCE":
site = node.attributes['VALUE'].value.strip().lower()
genericIntArgs = hostInfo.getElementsByTagName("GENERIC_VAR_INT")
for node in genericIntArgs:
if node.attributes['NAME'].value =="PILOT_" + username + "_VAR_5":
whetstones = int(node.attributes['VALUE'].value.strip().lower())
if whetstones > 65534:
whetstones = 0
# whetstones = 0
#if host is unknown, create a profiling task
currentHost = self.getHost(hostName)
if currentHost == None:
print ("Host/Pilot not found. hostname: " + hostName + ", LRMS: " + foundLrms)
if foundLrms == "jobmanager-pilot":
#he encontrado un pilot:
#primero busco e resource, y si no existe lo creo.
#luego creo un pilot que utilice ese resource
pilotResource = base.Session.query(PilotResource).filter(PilotResource.site == site, PilotResource.workerNode == workerNode).first()
if pilotResource == None:
print (" PilotResource was not found, creating a new one")
pilotResource = PilotResource(site, workerNode)
print (" Creating a new Pilot in NewPilotInfrastructure.createInfrastructureTasks")
newHost = Pilot(hostName, arch=foundArch, cpuMHz = foundCpuMHz, pilotResource = pilotResource, whetstones = whetstones)
self.pilots.append(newHost)
Session.add(newHost)
else:
print (" Creating a new Host in NewPilotInfrastructure.createInfrastructureTasks")
newHost = Host(hostName, arch=foundArch, cpuMHz = foundCpuMHz, lrms=foundLrms)
self.hosts.append(newHost)
Session.add(newHost)
#ESTO ES PARA HACER EL PROFILING DE LOS PILOT SI NO HAN PUBLICADO LOS WHETSTONES, SI NO NO HACE FALTA
#===============================================================
# if whetstones == 0 or whetstones > 65534:
# whetstones = 0
# print (" Host to profile: " + hostName + ": whetstone value not initialized ")
# hostsToProfile.append(newHost)
# #store new host on databae (faiulre resistance
# Session.add(newHost)
#===============================================================
#if information has changed, update host information
elif (currentHost.getWhetstones() != whetstones):
#va con un set porque es una operación más complicada, así que está encapsulada en esta funcion
currentHost.setWhetstones(whetstones)
Session.add(currentHost)
print ("Host: " + hostName + " UPDATED, new whetstones=" + str(whetstones))
elif currentHost.lrms == None:
currentHost.lrms = foundLrms
#pprofiling of new sites
hostProfilingTasks = [ExecutionManager.createHostProfilingTask(host)
for host in hostsToProfile
for i in range(base.profilingTasksPerHost)]
#estamos asumiento que todos los pilots publican la variable esa con su
#rendimiento, con lo que no hay que hacer el profiling de nada.
#AHORA, EN ESA NUEVA APROXIMACION, QUEREMOS TENER UNOS CUANTO SBENCHMARKS PARA IR ARRANCANDO PILOTS
print ("creating fake profiling tasks")
existingFakeTasks = len([task for task in infrastructureTasks if task.host.hostname=="" and task.status != "PENDING"])
existingGoodPilots = len (self.getGoodHosts())
existingProfilingTasks = len(hostProfilingTasks)
#fakeTasksToCreate = base.maxRunningTasks - (existingFakeTasks + existingGoodPilots + existingProfilingTasks)
fakeTasksToCreate = base.maxRunningTasks - existingFakeTasks
print (" Desired tasks: " + str(base.maxRunningTasks))
print (" Existing fake tasks: " + str(existingFakeTasks))
print (" Existing good pilots: " + str(existingGoodPilots))
print (" created: " + str(fakeTasksToCreate))
emptyHost = FakeHost()
fakeHostProfilingTasks = [ExecutionManager.createFakeHostProfilingTask(emptyHost)
for i in range(fakeTasksToCreate)]
hostProfilingTasks+=fakeHostProfilingTasks
return hostProfilingTasks
def load(self):
print ("LOAD - pilotInfrastructures")
#SITES
pastHosts = Session.query(Host).filter(Host.type=="hosts").all()
presentHosts = []
#PILOTS
pastPilots = Session.query(Pilot).all()
print ("Deleting all past pilots from database")
for pilot in pastPilots:
base.Session.delete(pilot)
presentPilots = []
GWHosts = obtainGWResources()
print ("Updating information")
#load XML to memory and extract data from hosts
try:
for resource in GWHosts:
hostname = resource.getElementsByTagName("HOSTNAME")[0].firstChild.data.strip().lower() #TODO: remove "unicode" from TEXT
try:
foundLrms = resource.getElementsByTagName("LRMS_NAME")[0].firstChild.data
except:
print ("Could not obtain resource LRMS for site " + hostname + ", skipping it")
continue
if foundLrms == "jobmanager-pilot":
#nombre de pilot
genericArgs = resource.getElementsByTagName("GENERIC_VAR_STR")
for node in genericArgs:
if node.attributes['NAME'].value =="PILOT_REAL_HOSTNAME":
workerNode = node.attributes['VALUE'].value.strip().lower()
if node.attributes['NAME'].value =="PILOT_REAL_RESOURCE":
site = node.attributes['VALUE'].value.strip().lower()
#numero de nodos libres. Esto en los pilots funciona bien, y es lo que se emplea para saber si está
#activo o apagado
try:
freeNodeCount = int(resource.getElementsByTagName("FREENODECOUNT")[0].firstChild.data)
except:
freeNodeCount = 0
if not freeNodeCount > 0:
continue;
#cargamos el pilot en la base de datos:
#primero buscamos el pilotResource, o lo creamos si no existe
#después creamos el pilot que emplee ese recurso
pilotResource = base.Session.query(PilotResource).filter(PilotResource.site == site, PilotResource.workerNode == workerNode).first()
if pilotResource == None:
pilotResource = PilotResource(site, workerNode)
pilot = Pilot(hostname, pilotResource=pilotResource)
presentPilots.append(pilot)
else: #it is a site
for host in pastHosts:
if host.hostname == hostname:
presentHosts.append(host)
break
except:
print("Error when parsing host information file, employing information from past executions")
self.hosts = pastHosts
self.pilots = []
return
#the hosts that we will first employ are the resulting ones
self.pilots = presentPilots
self.hosts = presentHosts
