def test_exec_dag_sink(self):
'''
Tests that the last function in a DAG executes correctly and stores the
result in the KVS.
'''
def func(_, x):
return x * x
fname = 'square'
arg = 2
dag = create_linear_dag([func], [fname], self.kvs_client, 'dag')
schedule, triggers = self._create_fn_schedule(dag, arg, fname, [fname])
exec_dag_function(self.pusher_cache, self.kvs_client, triggers, func,
schedule, self.user_library, {}, {})
# Assert that there have been 0 messages sent.
self.assertEqual(len(self.socket.outbox), 0)
# Retrieve the result, ensure it is a LWWPairLattice, then deserialize
# it.
result = self.kvs_client.get(schedule.id)[schedule.id]
self.assertEqual(type(result), LWWPairLattice)
result = serializer.load_lattice(result)
# Check that the output is equal to a local function execution.
self.assertEqual(result, func('', arg))
def test_exec_causal_dag_sink(self):
'''
Tests that the last function in a causal DAG executes correctly and
stores the result in the KVS. Also checks to make sure that causal
metadata is properly created.
'''
def func(_, x):
return x * x
fname = 'square'
arg = 2
dag = create_linear_dag([func], [fname], self.kvs_client, 'dag',
MultiKeyCausalLattice)
schedule, triggers = self._create_fn_schedule(dag, arg, fname, [fname],
MULTI)
schedule.output_key = 'output_key'
schedule.client_id = '12'
# We know that there is only one trigger. We populate dependencies
# explicitly in this trigger message to make sure that they are
# reflected in the final result.
kv = triggers['BEGIN'].dependencies.add()
kv.key = 'dependency'
DEFAULT_VC.serialize(kv.vector_clock)
exec_dag_function(self.pusher_cache, self.kvs_client, triggers, func,
schedule, self.user_library, {}, {})
# Assert that there have been 0 messages sent.
self.assertEqual(len(self.socket.outbox), 0)
# Retrieve the result and check its value and its metadata.
result = self.kvs_client.get(schedule.output_key)[schedule.output_key]
self.assertEqual(type(result), MultiKeyCausalLattice)
# Check that the vector clock of the output corresponds ot the client
# ID.
self.assertEqual(result.vector_clock,
VectorClock({schedule.client_id: 1}, True))
# Check that the dependencies of the output match those specified in
# the trigger.
self.assertEqual(len(result.dependencies.reveal()), 1)
self.assertTrue(kv.key in result.dependencies.reveal())
self.assertEqual(result.dependencies.reveal()[kv.key], DEFAULT_VC)
# Check that the output is equal to a local function execution.
result = serializer.load_lattice(result)[0]
self.assertEqual(result, func('', arg))
def test_exec_causal_dag_non_sink(self):
'''
Creates and executes a non-sink function in a causal-mode DAG. This
should be exactly the same as the non-causal version of the test,
except we ensure that the causal metadata is empty, because we don't
have any KVS accesses.
'''
# Create two functions intended to be used in sequence.
def incr(_, x):
x + 1
iname = 'incr'
def square(_, x):
return x * x
sname = 'square'
arg = 1
# Create a DAG and a trigger for the first function in the DAG.
dag = create_linear_dag([incr, square], [iname, sname],
self.kvs_client, 'dag', MultiKeyCausalLattice)
schedule, triggers = self._create_fn_schedule(dag, arg, iname,
[iname, sname], MULTI)
exec_dag_function(self.pusher_cache, self.kvs_client, triggers, incr,
schedule, self.user_library, {}, {})
# Assert that there has been a message sent.
self.assertEqual(len(self.pusher_cache.socket.outbox), 1)
# Extract that message and check its contents.
trigger = DagTrigger()
trigger.ParseFromString(self.pusher_cache.socket.outbox[0])
self.assertEqual(trigger.id, schedule.id)
self.assertEqual(trigger.target_function, sname)
self.assertEqual(trigger.source, iname)
self.assertEqual(len(trigger.arguments.values), 1)
self.assertEqual(len(trigger.version_locations), 0)
self.assertEqual(len(trigger.dependencies), 0)
val = serializer.load(trigger.arguments.values[0])
self.assertEqual(val, incr('', arg))
def test_exec_dag_non_sink(self):
'''
Executes a non-sink function in a DAG and ensures that the correct
downstream trigger was sent with a correct execution of the function.
'''
# Create two functions intended to be used in sequence.
def incr(_, x):
x + 1
iname = 'incr'
def square(_, x):
return x * x
sname = 'square'
arg = 1
# Create a DAG and a trigger for the first function in the DAG.
dag = create_linear_dag([incr, square], [iname, sname],
self.kvs_client, 'dag')
schedule, triggers = self._create_fn_schedule(dag, arg, iname,
[iname, sname])
exec_dag_function(self.pusher_cache, self.kvs_client, triggers, incr,
schedule, self.user_library, {}, {})
# Assert that there has been a message sent.
self.assertEqual(len(self.pusher_cache.socket.outbox), 1)
# Extract that message and check its contents.
trigger = DagTrigger()
trigger.ParseFromString(self.pusher_cache.socket.outbox[0])
self.assertEqual(trigger.id, schedule.id)
self.assertEqual(trigger.target_function, sname)
self.assertEqual(trigger.source, iname)
self.assertEqual(len(trigger.arguments.values), 1)
val = serializer.load(trigger.arguments.values[0])
self.assertEqual(val, incr('', arg))
def test_exec_causal_dag_non_sink_with_ref(self):
'''
Creates and executes a non-sink function in a causal-mode DAG. This
version accesses a KVS key, so we ensure that data is appropriately
cached and the metadata is passed downstream.
'''
# Create two functions intended to be used in sequence.
def incr(_, x):
x + 1
iname = 'incr'
def square(_, x):
return x * x
sname = 'square'
# Put tthe argument into the KVS.
arg_name = 'arg'
arg_value = 1
arg = serializer.dump_lattice(arg_value, MultiKeyCausalLattice)
self.kvs_client.put(arg_name, arg)
# Create a DAG and a trigger for the first function in the DAG.
dag = create_linear_dag([incr, square], [iname, sname],
self.kvs_client, 'dag', MultiKeyCausalLattice)
schedule, triggers = self._create_fn_schedule(
dag, DropletReference(arg_name, True), iname, [iname, sname],
MULTI)
exec_dag_function(self.pusher_cache, self.kvs_client, triggers, incr,
schedule, self.user_library, {}, {})
# Assert that there has been a message sent.
self.assertEqual(len(self.pusher_cache.socket.outbox), 1)
# Extract that message and check its contents.
trigger = DagTrigger()
trigger.ParseFromString(self.pusher_cache.socket.outbox[0])
self.assertEqual(trigger.id, schedule.id)
self.assertEqual(trigger.target_function, sname)
self.assertEqual(trigger.source, iname)
self.assertEqual(len(trigger.arguments.values), 1)
# Check the metadata of the key that is cached here after execution.
locs = trigger.version_locations
self.assertEqual(len(locs), 1)
self.assertTrue(self.ip in locs.keys())
self.assertEqual(len(locs[self.ip].keys), 1)
kv = locs[self.ip].keys[0]
self.assertEqual(kv.key, arg_name)
self.assertEqual(VectorClock(dict(kv.vector_clock), True),
arg.vector_clock)
# Check the metatada of the causal dependency passed downstream.
self.assertEqual(len(trigger.dependencies), 1)
kv = trigger.dependencies[0]
self.assertEqual(kv.key, arg_name)
self.assertEqual(VectorClock(dict(kv.vector_clock), True),
arg.vector_clock)
val = serializer.load(trigger.arguments.values[0])
self.assertEqual(val, incr('', arg_value))
def executor(ip, mgmt_ip, schedulers, thread_id):
logging.basicConfig(filename='log_executor.txt', level=logging.INFO,
format='%(asctime)s %(message)s')
context = zmq.Context(1)
poller = zmq.Poller()
pin_socket = context.socket(zmq.PULL)
pin_socket.bind(sutils.BIND_ADDR_TEMPLATE % (sutils.PIN_PORT + thread_id))
unpin_socket = context.socket(zmq.PULL)
unpin_socket.bind(sutils.BIND_ADDR_TEMPLATE % (sutils.UNPIN_PORT +
thread_id))
exec_socket = context.socket(zmq.PULL)
exec_socket.bind(sutils.BIND_ADDR_TEMPLATE % (sutils.FUNC_EXEC_PORT +
thread_id))
dag_queue_socket = context.socket(zmq.PULL)
dag_queue_socket.bind(sutils.BIND_ADDR_TEMPLATE % (sutils.DAG_QUEUE_PORT
+ thread_id))
dag_exec_socket = context.socket(zmq.PULL)
dag_exec_socket.bind(sutils.BIND_ADDR_TEMPLATE % (sutils.DAG_EXEC_PORT
+ thread_id))
self_depart_socket = context.socket(zmq.PULL)
self_depart_socket.bind(sutils.BIND_ADDR_TEMPLATE %
(sutils.SELF_DEPART_PORT + thread_id))
pusher_cache = SocketCache(context, zmq.PUSH)
poller = zmq.Poller()
poller.register(pin_socket, zmq.POLLIN)
poller.register(unpin_socket, zmq.POLLIN)
poller.register(exec_socket, zmq.POLLIN)
poller.register(dag_queue_socket, zmq.POLLIN)
poller.register(dag_exec_socket, zmq.POLLIN)
poller.register(self_depart_socket, zmq.POLLIN)
# If the management IP is set to None, that means that we are running in
# local mode, so we use a regular AnnaTcpClient rather than an IPC client.
if mgmt_ip:
client = AnnaIpcClient(thread_id, context)
else:
client = AnnaTcpClient('127.0.0.1', '127.0.0.1', local=True, offset=1)
user_library = DropletUserLibrary(context, pusher_cache, ip, thread_id,
client)
status = ThreadStatus()
status.ip = ip
status.tid = thread_id
status.running = True
utils.push_status(schedulers, pusher_cache, status)
departing = False
# Maintains a request queue for each function pinned on this executor. Each
# function will have a set of request IDs mapped to it, and this map stores
# a schedule for each request ID.
queue = {}
# Tracks the actual function objects that are pinned to this executor.
pinned_functions = {}
# Tracks runtime cost of excuting a DAG function.
runtimes = {}
# If multiple triggers are necessary for a function, track the triggers as
# we receive them. This is also used if a trigger arrives before its
# corresponding schedule.
received_triggers = {}
# Tracks when we received a function request, so we can report end-to-end
# latency for the whole executio.
receive_times = {}
# Tracks the number of requests we are finishing for each function pinned
# here.
exec_counts = {}
# Tracks the end-to-end runtime of each DAG request for which we are the
# sink function.
dag_runtimes = {}
# A map with KVS keys and their corresponding deserialized payloads.
cache = {}
# Internal metadata to track thread utilization.
report_start = time.time()
event_occupancy = {'pin': 0.0,
'unpin': 0.0,
'func_exec': 0.0,
'dag_queue': 0.0,
'dag_exec': 0.0}
total_occupancy = 0.0
while True:
socks = dict(poller.poll(timeout=1000))
if pin_socket in socks and socks[pin_socket] == zmq.POLLIN:
work_start = time.time()
pin(pin_socket, pusher_cache, client, status, pinned_functions,
runtimes, exec_counts)
utils.push_status(schedulers, pusher_cache, status)
elapsed = time.time() - work_start
event_occupancy['pin'] += elapsed
total_occupancy += elapsed
if unpin_socket in socks and socks[unpin_socket] == zmq.POLLIN:
work_start = time.time()
unpin(unpin_socket, status, pinned_functions, runtimes,
exec_counts)
utils.push_status(schedulers, pusher_cache, status)
elapsed = time.time() - work_start
event_occupancy['unpin'] += elapsed
total_occupancy += elapsed
if exec_socket in socks and socks[exec_socket] == zmq.POLLIN:
work_start = time.time()
exec_function(exec_socket, client, user_library, cache)
user_library.close()
utils.push_status(schedulers, pusher_cache, status)
elapsed = time.time() - work_start
event_occupancy['func_exec'] += elapsed
total_occupancy += elapsed
if dag_queue_socket in socks and socks[dag_queue_socket] == zmq.POLLIN:
work_start = time.time()
schedule = DagSchedule()
schedule.ParseFromString(dag_queue_socket.recv())
fname = schedule.target_function
logging.info('Received a schedule for DAG %s (%s), function %s.' %
(schedule.dag.name, schedule.id, fname))
if fname not in queue:
queue[fname] = {}
queue[fname][schedule.id] = schedule
if (schedule.id, fname) not in receive_times:
receive_times[(schedule.id, fname)] = time.time()
# In case we receive the trigger before we receive the schedule, we
# can trigger from this operation as well.
trkey = (schedule.id, fname)
if (trkey in received_triggers and (len(received_triggers[trkey])
== len(schedule.triggers))):
exec_dag_function(pusher_cache, client,
received_triggers[trkey],
pinned_functions[fname], schedule,
user_library, dag_runtimes, cache)
user_library.close()
del received_triggers[trkey]
del queue[fname][schedule.id]
fend = time.time()
fstart = receive_times[(schedule.id, fname)]
runtimes[fname].append(fend - fstart)
exec_counts[fname] += 1
elapsed = time.time() - work_start
event_occupancy['dag_queue'] += elapsed
total_occupancy += elapsed
if dag_exec_socket in socks and socks[dag_exec_socket] == zmq.POLLIN:
work_start = time.time()
trigger = DagTrigger()
trigger.ParseFromString(dag_exec_socket.recv())
fname = trigger.target_function
logging.info('Received a trigger for schedule %s, function %s.' %
(trigger.id, fname))
key = (trigger.id, fname)
if key not in received_triggers:
received_triggers[key] = {}
if (trigger.id, fname) not in receive_times:
receive_times[(trigger.id, fname)] = time.time()
received_triggers[key][trigger.source] = trigger
if fname in queue and trigger.id in queue[fname]:
schedule = queue[fname][trigger.id]
if len(received_triggers[key]) == len(schedule.triggers):
exec_dag_function(pusher_cache, client,
received_triggers[key],
pinned_functions[fname], schedule,
user_library, dag_runtimes, cache)
user_library.close()
del received_triggers[key]
del queue[fname][trigger.id]
fend = time.time()
fstart = receive_times[(trigger.id, fname)]
runtimes[fname].append(fend - fstart)
exec_counts[fname] += 1
elapsed = time.time() - work_start
event_occupancy['dag_exec'] += elapsed
total_occupancy += elapsed
if self_depart_socket in socks and socks[self_depart_socket] == \
zmq.POLLIN:
# This message does not matter.
self_depart_socket.recv()
logging.info('Preparing to depart. No longer accepting requests ' +
'and clearing all queues.')
status.ClearField('functions')
status.running = False
utils.push_status(schedulers, pusher_cache, status)
departing = True
# periodically report function occupancy
report_end = time.time()
if report_end - report_start > REPORT_THRESH:
cache.clear()
utilization = total_occupancy / (report_end - report_start)
status.utilization = utilization
# Periodically report my status to schedulers with the utilization
# set.
utils.push_status(schedulers, pusher_cache, status)
logging.info('Total thread occupancy: %.6f' % (utilization))
for event in event_occupancy:
occ = event_occupancy[event] / (report_end - report_start)
logging.info('\tEvent %s occupancy: %.6f' % (event, occ))
event_occupancy[event] = 0.0
stats = ExecutorStatistics()
for fname in runtimes:
if exec_counts[fname] > 0:
fstats = stats.functions.add()
fstats.name = fname
fstats.call_count = exec_counts[fname]
fstats.runtime.extend(runtimes[fname])
runtimes[fname].clear()
exec_counts[fname] = 0
for dname in dag_runtimes:
dstats = stats.dags.add()
dstats.name = dname
dstats.runtimes.extend(dag_runtimes[dname])
dag_runtimes[dname].clear()
# If we are running in cluster mode, mgmt_ip will be set, and we
# will report our status and statistics to it. Otherwise, we will
# write to the local conf file
if mgmt_ip:
sckt = pusher_cache.get(sutils.get_statistics_report_address
(mgmt_ip))
sckt.send(stats.SerializeToString())
sckt = pusher_cache.get(utils.get_util_report_address(mgmt_ip))
sckt.send(status.SerializeToString())
else:
logging.info(stats)
status.ClearField('utilization')
report_start = time.time()
total_occupancy = 0.0
# Periodically clear any old functions we have cached that we are
# no longer accepting requests for.
for fname in queue:
if len(queue[fname]) == 0 and fname not in status.functions:
del queue[fname]
del pinned_functions[fname]
del runtimes[fname]
del exec_counts[fname]
# If we are departing and have cleared our queues, let the
# management server know, and exit the process.
if departing and len(queue) == 0:
sckt = pusher_cache.get(utils.get_depart_done_addr(mgmt_ip))
sckt.send_string(ip)
# We specifically pass 1 as the exit code when ending our
# process so that the wrapper script does not restart us.
os._exit(1)