class RockerBoard(object):
"""
RockerBoard is a system balancing algorithm using popularity metrics to predict popularity
and make appropriate replications to keep the system balanced
"""
def __init__(self, config=dict()):
self.logger = logging.getLogger(__name__)
self.config = config
self.phedex = PhEDExService(self.config)
self.mit_db = MITDBService(self.config)
self.datasets = DatasetManager(self.config)
self.sites = SiteManager(self.config)
self.popularity = PopularityManager(self.config)
self.storage = StorageManager(self.config)
self.rankings = Ranker(self.config)
self.max_gb = int(self.config["rocker_board"]["max_gb"])
self.csv_data = list()
def start(self, date=datetime_day(datetime.datetime.utcnow())):
"""
Begin Rocker Board Algorithm
"""
t1 = datetime.datetime.utcnow()
# Get goals
dataset_rankings = self.rankings.get_dataset_rankings(date)
site_rankings = self.rankings.get_site_rankings(date)
self.change_dataset_rankings(dataset_rankings)
subscriptions = self.replicate(dataset_rankings, site_rankings)
self.logger.info("SUBSCRIPTIONS")
for subscription in subscriptions:
self.logger.info("site: %s\tdataset: %s", subscription[1], subscription[0])
# site_storage = self.rankings.get_site_storage_rankings(subscriptions)
# deletions = self.clean(dataset_rankings, site_storage)
# self.logger.info('DELETIONS')
# for deletion in deletions:
# self.logger.info('site: %s\tdataset: %s', deletion[1], deletion[0])
# self.delete(deletions)
self.subscribe(subscriptions)
# self.datasets.update_replicas(subscriptions, deletions)
t2 = datetime.datetime.utcnow()
td = t2 - t1
self.logger.info("Rocker Board took %s", str(td))
def change_dataset_rankings(self, dataset_rankings):
"""
Change the ranks from being the target number of replicas to being the
change in number of replicas required to reach the goal
"""
current_replicas = self.datasets.get_current_num_replicas()
for dataset in current_replicas:
dataset_rankings[dataset["name"]] -= dataset["n_replicas"]
def replicate(self, dataset_rankings, site_rankings):
"""
Balance system by creating new replicas based on popularity
"""
subscriptions = list()
subscribed_gb = 0
sites_available_storage_gb = self.sites.get_all_available_storage()
while (subscribed_gb < self.max_gb) and site_rankings:
tmp_site_rankings = dict()
for k, v in site_rankings.items():
tmp_site_rankings[k] = v
dataset = max(dataset_rankings.iteritems(), key=operator.itemgetter(1))
dataset_name = dataset[0]
dataset_rank = dataset[1]
if (not dataset_name) or (dataset_rank < 1):
break
size_gb = self.datasets.get_size(dataset_name)
unavailable_sites = set(self.datasets.get_sites(dataset_name))
for site_name in tmp_site_rankings.keys():
if (self.sites.get_available_storage(site_name) < size_gb) or (tmp_site_rankings[site_name] <= 0):
unavailable_sites.add(site_name)
for site_name in unavailable_sites:
try:
del tmp_site_rankings[site_name]
except:
continue
if not tmp_site_rankings:
del dataset_rankings[dataset_name]
continue
site_name = weighted_choice(tmp_site_rankings)
subscription = (dataset_name, site_name)
subscriptions.append(subscription)
subscribed_gb += size_gb
sites_available_storage_gb[site_name] -= size_gb
self.logger.info("%s : added", dataset_name)
if sites_available_storage_gb[site_name] <= 0:
del site_rankings[site_name]
dataset_rankings[dataset_name] -= 1
self.logger.info("Subscribed %dGB", subscribed_gb)
return subscriptions
def clean(self, dataset_rankings, site_rankings):
"""
Suggest deletions based on dataset and site rankings
"""
deletions = list()
deleted_gb = 0
while site_rankings:
tmp_site_rankings = dict()
dataset = min(dataset_rankings.iteritems(), key=operator.itemgetter(1))
dataset_name = dataset[0]
size_gb = self.datasets.get_size(dataset_name)
available_sites = set(self.datasets.get_sites(dataset_name))
for site_name in available_sites:
try:
tmp_site_rankings[site_name] = site_rankings[site_name]
except:
continue
if not tmp_site_rankings:
del dataset_rankings[dataset_name]
continue
site_name = weighted_choice(tmp_site_rankings)
deletion = (dataset_name, site_name)
deletions.append(deletion)
deleted_gb += size_gb
site_rankings[site_name] -= size_gb
dataset_rankings[dataset_name] += 1
if site_rankings[site_name] <= 0:
del site_rankings[site_name]
self.logger.info("Deleted %dGB", deleted_gb)
return deletions
def subscribe(self, subscriptions):
"""
Make subscriptions to phedex
subscriptions = [(dataset_name, site_name), ...]
"""
new_subscriptions = dict()
for subscription in subscriptions:
dataset_name = subscription[0]
site_name = subscription[1]
try:
new_subscriptions[site_name].append(dataset_name)
except:
new_subscriptions[site_name] = list()
new_subscriptions[site_name].append(dataset_name)
for site_name, dataset_names in new_subscriptions.items():
data = self.phedex.generate_xml(dataset_names)
comments = (
"This dataset is predicted to become popular and has therefore been automatically replicated by cuadrnt"
)
api = "subscribe"
params = [
("node", site_name),
("data", data),
("level", "dataset"),
("move", "n"),
("custodial", "n"),
("group", "AnalysisOps"),
("request_only", "n"),
("no_mail", "n"),
("comments", comments),
]
json_data = self.phedex.fetch(api=api, params=params, method="post")
# insert into db
group_name = "AnalysisOps"
request_id = 0
request_type = 0
try:
request = json_data["phedex"]
request_id = request["request_created"][0]["id"]
request_created = timestamp_to_datetime(request["request_timestamp"])
except:
self.logger.warning(
"Subscription did not succeed\n\tSite:%s\n\tDatasets: %s", str(site_name), str(dataset_names)
)
continue
for dataset_name in dataset_names:
coll = "dataset_rankings"
date = datetime_day(datetime.datetime.utcnow())
pipeline = list()
match = {"$match": {"name": dataset_name, "date": date}}
pipeline.append(match)
project = {"$project": {"delta_rank": 1, "_id": 0}}
pipeline.append(project)
data = self.storage.get_data(coll=coll, pipeline=pipeline)
dataset_rank = data[0]["delta_rank"]
query = "INSERT INTO Requests(RequestId, RequestType, DatasetId, SiteId, GroupId, Rank, Date) SELECT %s, %s, Datasets.DatasetId, Sites.SiteId, Groups.GroupId, %s, %s FROM Datasets, Sites, Groups WHERE Datasets.DatasetName=%s AND Sites.SiteName=%s AND Groups.GroupName=%s"
values = (request_id, request_type, dataset_rank, request_created, dataset_name, site_name, group_name)
self.mit_db.query(query=query, values=values, cache=False)
def delete(self, deletions):
"""
Make deletions to phedex
deletions = [(dataset_name, site_name), ...]
"""
new_deletions = dict()
for deletion in deletions:
dataset_name = deletion[0]
site_name = deletion[1]
try:
new_deletions[site_name].append(dataset_name)
except:
new_deletions[site_name] = list()
new_deletions[site_name].append(dataset_name)
for site_name, dataset_names in new_deletions.items():
data = self.phedex.generate_xml(dataset_names)
comments = "This dataset is predicted to become less popular and has therefore been automatically deleted by cuadrnt"
api = "delete"
params = [
("node", site_name),
("data", data),
("level", "dataset"),
("rm_subscriptions", "y"),
("comments", comments),
]
json_data = self.phedex.fetch(api=api, params=params, method="post")
# insert into db
group_name = "AnalysisOps"
request_id = 0
request_type = 1
try:
request = json_data["phedex"]
request_id = request["request_created"][0]["id"]
request_created = timestamp_to_datetime(request["request_timestamp"])
except:
self.logger.warning(
"Deletion did not succeed\n\tSite:%s\n\tDatasets: %s", str(site_name), str(dataset_names)
)
continue
for dataset_name in dataset_names:
coll = "dataset_rankings"
date = datetime_day(datetime.datetime.utcnow())
pipeline = list()
match = {"$match": {"name": dataset_name, "date": date}}
pipeline.append(match)
project = {"$project": {"delta_rank": 1, "_id": 0}}
pipeline.append(project)
data = self.storage.get_data(coll=coll, pipeline=pipeline)
dataset_rank = data[0]["delta_rank"]
query = "INSERT INTO Requests(RequestId, RequestType, DatasetId, SiteId, GroupId, Rank, Date) SELECT %s, %s, Datasets.DatasetId, Sites.SiteId, Groups.GroupId, %s, %s FROM Datasets, Sites, Groups WHERE Datasets.DatasetName=%s AND Sites.SiteName=%s AND Groups.GroupName=%s"
values = (request_id, request_type, dataset_rank, request_created, dataset_name, site_name, group_name)
self.mit_db.query(query=query, values=values, cache=False)
class Ranker(object):
"""
Generic Ranking class
"""
def __init__(self, config=dict()):
self.logger = logging.getLogger(__name__)
self.config = config
self.sites = SiteManager(self.config)
self.datasets = DatasetManager(self.config)
self.popularity = PopularityManager(self.config)
self.storage = StorageManager(self.config)
self.max_replicas = int(config['rocker_board']['max_replicas'])
self.MAX_THREADS = int(config['threading']['max_threads'])
self.dataset_popularity = dict()
def get_dataset_rankings(self, date=datetime_day(datetime.datetime.utcnow())):
"""
Generate dataset rankings
"""
self.dataset_popularity = dict()
dataset_names = self.datasets.get_db_datasets()
q = Queue.Queue()
for i in range(self.MAX_THREADS):
worker = threading.Thread(target=self.get_dataset_popularity, args=(q,))
worker.daemon = True
worker.start()
# self.dataset_features = self.popularity.get_features(dataset_names, date)
# self.dataset_tiers = self.datasets.get_data_tiers(dataset_names)
for dataset_name in dataset_names:
q.put((dataset_name, date))
q.join()
dataset_rankings = self.normalize_popularity(date)
return dataset_rankings
def get_site_rankings(self, date=datetime_day(datetime.datetime.utcnow())):
"""
Generate site rankings
"""
# get all sites which can be replicated to
site_names = self.sites.get_available_sites()
site_rankings = dict()
for site_name in site_names:
# get popularity
popularity = self.get_site_popularity(site_name, date)
# get cpu and storage (performance)
performance = self.sites.get_performance(site_name)
# get available storage
available_storage_tb = self.sites.get_available_storage(site_name)/10**3
if available_storage_tb <= 0:
available_storage_tb = 0
else:
available_storage_tb = 1
#calculate rank
try:
rank = (performance*available_storage_tb)/popularity
except:
rank = 0.0
# store into dict
site_rankings[site_name] = rank
# insert into database
coll = 'site_rankings'
query = {'name':site_name, 'date':date}
data = {'$set':{'name':site_name, 'date':date, 'rank':rank, 'popularity':popularity}}
self.storage.update_data(coll=coll, query=query, data=data, upsert=True)
return site_rankings
def get_dataset_popularity(self, q):
"""
Get the estimated popularity for dataset
"""
while True:
# collect features
data = q.get()
dataset_name = data[0]
date = data[1]
popularity = 0.0
# get average
popularity = self.popularity.get_average_popularity(dataset_name, date)
self.dataset_popularity[dataset_name] = popularity
q.task_done()
def get_site_popularity(self, site_name, date=datetime_day(datetime.datetime.utcnow())):
"""
Get popularity for site
"""
# get all datasets with a replica at the site and how many replicas it has
coll = 'dataset_data'
pipeline = list()
match = {'$match':{'replicas':site_name}}
pipeline.append(match)
project = {'$project':{'name':1, '_id':0}}
pipeline.append(project)
data = self.storage.get_data(coll=coll, pipeline=pipeline)
popularity = 0.0
dataset_names = [dataset_data['name'] for dataset_data in data]
# get the popularity of the dataset and decide by number of replicas
coll = 'dataset_rankings'
pipeline = list()
match = {'$match':{'date':date}}
pipeline.append(match)
match = {'$match':{'name':{'$in':dataset_names}}}
pipeline.append(match)
group = {'$group':{'_id':'$date', 'total_popularity':{'$sum':'$popularity'}}}
pipeline.append(group)
project = {'$project':{'total_popularity':1, '_id':0}}
pipeline.append(project)
data = self.storage.get_data(coll=coll, pipeline=pipeline)
try:
popularity = data[0]['total_popularity']
except:
popularity = 0.0
return popularity
def get_site_storage_rankings(self, subscriptions):
"""
Return the amount over the soft limit sites are including new subscriptions
If site is not over just set to 0
"""
site_rankings = dict()
available_sites = self.sites.get_available_sites()
for site_name in available_sites:
site_rankings[site_name] = self.sites.get_over_soft_limit(site_name)
for subscription in subscriptions:
site_rankings[subscription[1]] += self.datasets.get_size(subscription[0])
for site_name in available_sites:
if site_rankings[site_name] < 0:
del site_rankings[site_name]
return site_rankings
def normalize_popularity(self, date):
"""
Normalize popularity values to be between 1 and max_replicas
"""
dataset_rankings = dict()
max_pop = max(self.dataset_popularity.iteritems(), key=operator.itemgetter(1))[1]
min_pop = min(self.dataset_popularity.iteritems(), key=operator.itemgetter(1))[1]
n = float(min_pop + (self.max_replicas - 1))/max_pop
m = 1 - n*min_pop
for dataset_name, popularity in self.dataset_popularity.items():
# store into dict
rank = int(n*self.dataset_popularity[dataset_name] + m)
dataset_rankings[dataset_name] = rank
coll = 'dataset_rankings'
query = data = {'name':dataset_name, 'date':date}
data = {'$set':{'name':dataset_name, 'date':date, 'rank':rank, 'popularity':popularity}}
self.storage.update_data(coll=coll, query=query, data=data, upsert=True)
return dataset_rankings
class Ranker(object):
"""
Generic Ranking class
"""
def __init__(self, config=dict()):
self.logger = logging.getLogger(__name__)
self.config = config
self.sites = SiteManager(self.config)
self.datasets = DatasetManager(self.config)
self.popularity = PopularityManager(self.config)
self.storage = StorageManager(self.config)
self.max_replicas = int(config['rocker_board']['max_replicas'])
self.MAX_THREADS = int(config['threading']['max_threads'])
self.dataset_popularity = dict()
def get_dataset_rankings(self,
date=datetime_day(datetime.datetime.utcnow())):
"""
Generate dataset rankings
"""
self.dataset_popularity = dict()
dataset_names = self.datasets.get_db_datasets()
q = Queue.Queue()
for i in range(self.MAX_THREADS):
worker = threading.Thread(target=self.get_dataset_popularity,
args=(q, ))
worker.daemon = True
worker.start()
# self.dataset_features = self.popularity.get_features(dataset_names, date)
# self.dataset_tiers = self.datasets.get_data_tiers(dataset_names)
for dataset_name in dataset_names:
q.put((dataset_name, date))
q.join()
dataset_rankings = self.normalize_popularity(date)
return dataset_rankings
def get_site_rankings(self, date=datetime_day(datetime.datetime.utcnow())):
"""
Generate site rankings
"""
# get all sites which can be replicated to
site_names = self.sites.get_available_sites()
site_rankings = dict()
for site_name in site_names:
# get popularity
popularity = self.get_site_popularity(site_name, date)
# get cpu and storage (performance)
performance = self.sites.get_performance(site_name)
# get available storage
available_storage_tb = self.sites.get_available_storage(
site_name) / 10**3
if available_storage_tb <= 0:
available_storage_tb = 0
else:
available_storage_tb = 1
#calculate rank
try:
rank = (performance * available_storage_tb) / popularity
except:
rank = 0.0
# store into dict
site_rankings[site_name] = rank
# insert into database
coll = 'site_rankings'
query = {'name': site_name, 'date': date}
data = {
'$set': {
'name': site_name,
'date': date,
'rank': rank,
'popularity': popularity
}
}
self.storage.update_data(coll=coll,
query=query,
data=data,
upsert=True)
return site_rankings
def get_dataset_popularity(self, q):
"""
Get the estimated popularity for dataset
"""
while True:
# collect features
data = q.get()
dataset_name = data[0]
date = data[1]
popularity = 0.0
# get average
popularity = self.popularity.get_average_popularity(
dataset_name, date)
self.dataset_popularity[dataset_name] = popularity
q.task_done()
def get_site_popularity(self,
site_name,
date=datetime_day(datetime.datetime.utcnow())):
"""
Get popularity for site
"""
# get all datasets with a replica at the site and how many replicas it has
coll = 'dataset_data'
pipeline = list()
match = {'$match': {'replicas': site_name}}
pipeline.append(match)
project = {'$project': {'name': 1, '_id': 0}}
pipeline.append(project)
data = self.storage.get_data(coll=coll, pipeline=pipeline)
popularity = 0.0
dataset_names = [dataset_data['name'] for dataset_data in data]
# get the popularity of the dataset and decide by number of replicas
coll = 'dataset_rankings'
pipeline = list()
match = {'$match': {'date': date}}
pipeline.append(match)
match = {'$match': {'name': {'$in': dataset_names}}}
pipeline.append(match)
group = {
'$group': {
'_id': '$date',
'total_popularity': {
'$sum': '$popularity'
}
}
}
pipeline.append(group)
project = {'$project': {'total_popularity': 1, '_id': 0}}
pipeline.append(project)
data = self.storage.get_data(coll=coll, pipeline=pipeline)
try:
popularity = data[0]['total_popularity']
except:
popularity = 0.0
return popularity
def get_site_storage_rankings(self, subscriptions):
"""
Return the amount over the soft limit sites are including new subscriptions
If site is not over just set to 0
"""
site_rankings = dict()
available_sites = self.sites.get_available_sites()
for site_name in available_sites:
site_rankings[site_name] = self.sites.get_over_soft_limit(
site_name)
for subscription in subscriptions:
site_rankings[subscription[1]] += self.datasets.get_size(
subscription[0])
for site_name in available_sites:
if site_rankings[site_name] < 0:
del site_rankings[site_name]
return site_rankings
def normalize_popularity(self, date):
"""
Normalize popularity values to be between 1 and max_replicas
"""
dataset_rankings = dict()
max_pop = max(self.dataset_popularity.iteritems(),
key=operator.itemgetter(1))[1]
min_pop = min(self.dataset_popularity.iteritems(),
key=operator.itemgetter(1))[1]
n = float(min_pop + (self.max_replicas - 1)) / max_pop
m = 1 - n * min_pop
for dataset_name, popularity in self.dataset_popularity.items():
# store into dict
rank = int(n * self.dataset_popularity[dataset_name] + m)
dataset_rankings[dataset_name] = rank
coll = 'dataset_rankings'
query = data = {'name': dataset_name, 'date': date}
data = {
'$set': {
'name': dataset_name,
'date': date,
'rank': rank,
'popularity': popularity
}
}
self.storage.update_data(coll=coll,
query=query,
data=data,
upsert=True)
return dataset_rankings
class RockerBoard(object):
"""
RockerBoard is a system balancing algorithm using popularity metrics to predict popularity
and make appropriate replications to keep the system balanced
"""
def __init__(self, config=dict()):
self.logger = logging.getLogger(__name__)
self.config = config
self.phedex = PhEDExService(self.config)
self.mit_db = MITDBService(self.config)
self.datasets = DatasetManager(self.config)
self.sites = SiteManager(self.config)
self.popularity = PopularityManager(self.config)
self.storage = StorageManager(self.config)
self.rankings = Ranker(self.config)
self.max_gb = int(self.config['rocker_board']['max_gb'])
self.csv_data = list()
def start(self, date=datetime_day(datetime.datetime.utcnow())):
"""
Begin Rocker Board Algorithm
"""
t1 = datetime.datetime.utcnow()
# Get goals
dataset_rankings = self.rankings.get_dataset_rankings(date)
site_rankings = self.rankings.get_site_rankings(date)
self.change_dataset_rankings(dataset_rankings)
subscriptions = self.replicate(dataset_rankings, site_rankings)
self.logger.info('SUBSCRIPTIONS')
for subscription in subscriptions:
self.logger.info('site: %s\tdataset: %s', subscription[1], subscription[0])
# self.subscribe(subscriptions)
t2 = datetime.datetime.utcnow()
td = t2 - t1
self.logger.info('Rocker Board took %s', str(td))
def change_dataset_rankings(self, dataset_rankings):
"""
Change the ranks from being the target number of replicas to being the
change in number of replicas required to reach the goal
"""
current_replicas = self.datasets.get_current_num_replicas()
for dataset in current_replicas:
dataset_rankings[dataset['name']] -= dataset['n_replicas']
def replicate(self, dataset_rankings, site_rankings):
"""
Balance system by creating new replicas based on popularity
"""
subscriptions = list()
subscribed_gb = 0
sites_available_storage_gb = self.sites.get_all_available_storage()
while (subscribed_gb < self.max_gb) and site_rankings:
tmp_site_rankings = dict()
for k, v in site_rankings.items():
tmp_site_rankings[k] = v
dataset = max(dataset_rankings.iteritems(), key=operator.itemgetter(1))
dataset_name = dataset[0]
dataset_rank = dataset[1]
if (not dataset_name) or (dataset_rank < 1):
break
size_gb = self.datasets.get_size(dataset_name)
unavailable_sites = set(self.datasets.get_sites(dataset_name))
for site_name in tmp_site_rankings.keys():
if (self.sites.get_available_storage(site_name) < size_gb) or (tmp_site_rankings[site_name] <= 0):
unavailable_sites.add(site_name)
for site_name in unavailable_sites:
try:
del tmp_site_rankings[site_name]
except:
continue
if not tmp_site_rankings:
del dataset_rankings[dataset_name]
continue
site_name = weighted_choice(tmp_site_rankings)
subscription = (dataset_name, site_name)
subscriptions.append(subscription)
subscribed_gb += size_gb
sites_available_storage_gb[site_name] -= size_gb
self.logger.info('%s : added', dataset_name)
if sites_available_storage_gb[site_name] <= 0:
del site_rankings[site_name]
dataset_rankings[dataset_name] -= 1
self.logger.info('Subscribed %dGB', subscribed_gb)
return subscriptions
def subscribe(self, subscriptions):
"""
Make subscriptions to phedex
subscriptions = [(dataset_name, site_name), ...]
"""
new_subscriptions = dict()
for subscription in subscriptions:
dataset_name = subscription[0]
site_name = subscription[1]
try:
new_subscriptions[site_name].append(dataset_name)
except:
new_subscriptions[site_name] = list()
new_subscriptions[site_name].append(dataset_name)
for site_name, dataset_names in new_subscriptions.items():
data = self.phedex.generate_xml(dataset_names)
comments = 'This dataset is predicted to become popular and has therefore been automatically replicated by cuadrnt'
api = 'subscribe'
params = [('node', site_name), ('data', data), ('level','dataset'), ('move', 'n'), ('custodial', 'n'), ('group', 'AnalysisOps'), ('request_only', 'n'), ('no_mail', 'n'), ('comments', comments)]
json_data = self.phedex.fetch(api=api, params=params, method='post')
# insert into db
group_name = 'AnalysisOps'
request_id = 0
request_type = 0
try:
request = json_data['phedex']
request_id = request['request_created'][0]['id']
request_created = timestamp_to_datetime(request['request_timestamp'])
except:
self.logger.warning('Subscription did not succeed\n\tSite:%s\n\tDatasets: %s', str(site_name), str(dataset_names))
continue
for dataset_name in dataset_names:
coll = 'dataset_rankings'
date = datetime_day(datetime.datetime.utcnow())
pipeline = list()
match = {'$match':{'name':dataset_name, 'date':date}}
pipeline.append(match)
project = {'$project':{'delta_rank':1, '_id':0}}
pipeline.append(project)
data = self.storage.get_data(coll=coll, pipeline=pipeline)
dataset_rank = data[0]['delta_rank']
query = "INSERT INTO Requests(RequestId, RequestType, DatasetId, SiteId, GroupId, Rank, Date) SELECT %s, %s, Datasets.DatasetId, Sites.SiteId, Groups.GroupId, %s, %s FROM Datasets, Sites, Groups WHERE Datasets.DatasetName=%s AND Sites.SiteName=%s AND Groups.GroupName=%s"
values = (request_id, request_type, dataset_rank, request_created, dataset_name, site_name, group_name)
self.mit_db.query(query=query, values=values, cache=False)