def __init__(self):
# observation and action space
self.setup_space()
# random seed
self.seed(config.seed)
# global timer
self.wall_time = WallTime()
# uses priority queue
self.timeline = Timeline()
# executors
self.executors = OrderedSet()
for exec_id in range(config.exec_cap):
self.executors.add(Executor(exec_id))
# free executors
self.free_executors = FreeExecutors(self.executors)
# moving executors
self.moving_executors = MovingExecutors()
# executor commit
self.exec_commit = ExecutorCommit()
# prevent agent keeps selecting the same node
self.node_selected = set()
# for computing reward at each step
self.reward_calculator = RewardCalculator()
def reset(self, max_time=np.inf):
# reset observation and action space
self.setup_space()
self.max_time = max_time
self.wall_time.reset()
self.timeline.reset()
self.exec_commit.reset()
self.moving_executors.reset()
self.reward_calculator.reset()
self.finished_job_dags = OrderedSet()
self.node_selected.clear()
for executor in self.executors:
executor.reset()
self.free_executors.reset(self.executors)
# generate a set of new jobs
self.job_dags = generate_jobs(
self.np_random, self.timeline, self.wall_time)
# map action to dag_idx and node_idx
self.action_map = compute_act_map(self.job_dags)
# add initial set of jobs in the system
for job_dag in self.job_dags:
self.add_job(job_dag)
# put all executors as source executors initially
self.source_job = None
self.num_source_exec = len(self.executors)
self.exec_to_schedule = OrderedSet(self.executors)
return self.observe()
def __init__(self, dag_db):
self.dag_db = dag_db
self.job_dags = OrderedSet()
self.action_map = {} # action index -> node
self.available_executors = {}
self.last_trigger = None
# executors
self.executors = {}
for exec_id in range(config.exec_cap):
self.executors[exec_id] = Executor(exec_id)
# dynamically bind {app_id -> job_dag}
self.spark_dag_map = {}
# dynamically bind {job_dag -> app_id}
self.spark_inverse_dag_map = {}
# dynamically bind {(app_id, stage_id) -> node}
self.spark_node_map = {}
# dynamically bind {node -> (app_id, stage_id)}
self.spark_inverse_node_map = {}
# dynamically bind {app_id -> {exec_id -> re-usable track_id}}
self.exec_id_track_id_map = {}
def generate_jobs(np_random, timeline, wall_time):
job_dags = OrderedSet()
tpch_size = ['2g','5g','10g','20g','50g','80g','100g']
tpch_num = 22
t = 0
for _ in range(config.num_init_dags):
# generate query
query_size = tpch_size[np_random.randint(len(tpch_size))]
query_idx = str(np_random.randint(tpch_num) + 1)
# generate job
job_dag = load_job(
query_size, query_idx, wall_time, np_random)
# job already arrived, put in job_dags
job_dag.start_time = t
job_dag.arrived = True
job_dags.add(job_dag)
for _ in range(config.num_stream_dags):
# poisson process
t += int(np_random.exponential(config.stream_interval))
# uniform distribution
query_size = tpch_size[np_random.randint(len(tpch_size))]
query_idx = str(np_random.randint(tpch_num) + 1)
# generate job
job_dag = load_job(
query_size, query_idx, wall_time, np_random)
# push into timeline
job_dag.start_time = t
timeline.push(t, job_dag)
return job_dags
def reset(self):
for node in self.nodes:
node.reset()
self.num_nodes_done = 0
self.executors = OrderedSet()
self.frontier_nodes = OrderedSet()
for node in self.nodes:
if node.is_schedulable():
self.frontier_nodes.add(node)
self.arrived = False
self.completed = False
self.completion_time = np.inf
def get_frontier_nodes(self):
# frontier nodes := unsaturated nodes with all parent nodes saturated
frontier_nodes = OrderedSet()
for job_dag in self.job_dags:
for node in job_dag.nodes:
if not node in self.node_selected and not self.saturated(node):
parents_saturated = True
for parent_node in node.parent_nodes:
if not self.saturated(parent_node):
parents_saturated = False
break
if parents_saturated:
frontier_nodes.add(node)
return frontier_nodes
def __init__(self, nodes, adj_mat, name):
# nodes: list of N nodes
# adj_mat: N by N 0-1 adjacency matrix, e_ij = 1 -> edge from i to j
assert len(nodes) == adj_mat.shape[0]
assert adj_mat.shape[0] == adj_mat.shape[1]
self.name = name
self.nodes = nodes
self.adj_mat = adj_mat
self.num_nodes = len(self.nodes)
self.num_nodes_done = 0
# set of executors currently running on the job
self.executors = OrderedSet()
# the computation graph needs to be a DAG
assert is_dag(self.num_nodes, self.adj_mat)
# get the set of schedule nodes
self.frontier_nodes = OrderedSet()
for node in self.nodes:
if node.is_schedulable():
self.frontier_nodes.add(node)
# assign job dag to node
self.assign_job_dag_to_node()
# dag is arrived
self.arrived = False
# dag is completed
self.completed = False
# dag start ime
self.start_time = None
# dag completion time
self.completion_time = np.inf
# map a executor number to an interval
self.executor_interval_map = \
self.get_executor_interval_map()
def assign_executor(self, executor, frontier_changed):
if executor.node is not None and not executor.node.no_more_tasks:
# keep working on the previous node
task = executor.node.schedule(executor)
self.timeline.push(task.finish_time, task)
else:
# need to move on to other nodes
if frontier_changed:
# frontier changed, need to consult all free executors
# note: executor.job_dag might change after self.schedule()
source_job = executor.job_dag
if len(self.exec_commit[executor.node]) > 0:
# directly fulfill the commitment
self.exec_to_schedule = {executor}
self.schedule()
else:
# free up the executor
self.free_executors.add(source_job, executor)
# then consult all free executors
self.exec_to_schedule = OrderedSet(self.free_executors[source_job])
self.source_job = source_job
self.num_source_exec = len(self.free_executors[source_job])
else:
# just need to schedule one current executor
self.exec_to_schedule = {executor}
# only care about executors on the node
if len(self.exec_commit[executor.node]) > 0:
# directly fulfill the commitment
self.schedule()
else:
# need to consult for ALL executors on the node
# Note: self.exec_to_schedule is immediate
# self.num_source_exec is for commit
# so len(self.exec_to_schedule) !=
# self.num_source_exec can happen
self.source_job = executor.job_dag
self.num_source_exec = len(executor.node.executors)
def __init__(self, idx, tasks, task_duration, wall_time, np_random):
self.idx = idx
self.tasks = tasks
self.wall_time = wall_time
self.np_random = np_random
self.task_duration = task_duration
self.num_tasks = len(tasks)
self.num_finished_tasks = 0
self.next_task_idx = 0
self.no_more_tasks = False
self.tasks_all_done = False
self.node_finish_time = np.inf
self.executors = OrderedSet()
# uninitialized
self.parent_nodes = []
self.child_nodes = []
self.descendant_nodes = []
self.job_dag = None
self.assign_node_to_tasks()
def run(self):
# set up ipc communication
context = zmq.Context()
socket = context.socket(zmq.REP)
ipc_msg = IPCMessage()
ipc_reply = IPCReply()
os.system('rm /tmp/spark_scheduling_java_python_ipc')
socket.bind("ipc:///tmp/spark_scheduling_java_python_ipc")
# for reward computation
num_active_jobs = 0
prev_time = time.time()
while not self.exit.is_set():
msg = socket.recv()
ipc_msg.ParseFromString(msg)
if ipc_msg.msg_type == 'register':
self.dag_db.add_new_app(ipc_msg.app_name, ipc_msg.app_id)
job_dag = self.env.add_job_dag(ipc_msg.app_id)
add_job_in_graph(self.graph, job_dag)
ipc_reply.msg = \
"external scheduler register app " + str(ipc_msg.app_name)
elif ipc_msg.msg_type == 'bind':
self.env.bind_exec_id(ipc_msg.app_id, ipc_msg.exec_id,
ipc_msg.track_id)
ipc_reply.msg = \
"external scheduler bind app_id " + \
str(ipc_msg.app_id) + " exec_id " + \
str(ipc_msg.exec_id) + " on track_id " + \
str(ipc_msg.track_id)
elif ipc_msg.msg_type == 'inform':
self.env.complete_tasks(ipc_msg.app_id, ipc_msg.stage_id,
ipc_msg.num_tasks_left)
ipc_reply.msg = \
"external scheduler updated app_id " + \
str(ipc_msg.app_id) + \
" stage_id " + \
str(ipc_msg.stage_id) + \
" with " + str(ipc_msg.num_tasks_left) + " tasks left"
elif ipc_msg.msg_type == 'update':
frontier_nodes_changed = \
self.env.complete_stage(ipc_msg.app_id, ipc_msg.stage_id)
ipc_reply.msg = \
"external scheduler updated app_id " + \
str(ipc_msg.app_id) + \
" stage_id " + \
str(ipc_msg.stage_id)
elif ipc_msg.msg_type == 'tracking':
# master asks which app it should assign the executor to
ipc_reply.app_id, ipc_reply.num_executors_to_take = \
self.exec_tracker.pop_executor_flow(ipc_msg.num_available_executors)
ipc_reply.msg = \
"external scheduler moves " + \
str(ipc_reply.num_executors_to_take) + \
" executor to app " + ipc_reply.app_id
elif ipc_msg.msg_type == 'consult':
# convert ipc_msg.app_id and ipc_msg.stage_id to corresponding
# executors in virtual environment and then inovke the
# scheduling agent
# 1. translate the raw information into observation space
# sort out the exec_map (where the executors are)
exec_map = {job_dag: 0 for job_dag in self.env.job_dags}
for app_id in self.dag_db.apps_map:
if app_id in self.exec_tracker.executor_flow:
job_dag = self.dag_db.apps_map[app_id]
exec_map[job_dag] = self.exec_tracker.executor_flow[
app_id]
source_job = self.dag_db.apps_map[ipc_msg.app_id]
frontier_nodes = OrderedSet()
for job_dag in self.env.job_dags:
for node in job_dag.frontier_nodes:
frontier_nodes.add(node)
for job_dag in self.env.job_dags:
for node in job_dag.nodes:
feature = np.zeros([6])
# number of executors already in the job
feature[0] = exec_map[job_dag]
# source executor is from the current job (locality)
feature[1] = job_dag is source_job
# number of source executors
feature[2] = 1
# remaining number of tasks in the node
feature[3] = node.num_tasks - node.next_task_idx
# average task duration of the node
feature[4] = node.tasks[-1].duration
# is the current node valid
feature[5] = node in frontier_nodes
# update feature in observation
self.graph.update_nodes({node: feature})
# update mask in the action space
self.action_space.update_valid_set(frontier_nodes)
# 2. gather feedback for the previous action
curr_time = time.time()
elapsed_time = curr_time - prev_time
prev_reward = num_active_jobs * elapsed_time
prev_done = False # spark can be long running
prev_info = {'elapsed_time': elapsed_time}
num_active_jobs = len(self.env.job_dags)
prev_time = curr_time
# 3. get the action from the agent
node = self.agent.get_action(self.graph, prev_reward,
prev_done, prev_info)
# 4. translate the action to ipc reply
if node is None:
# no-action was made
ipc_reply.app_id = 'void'
ipc_reply.stage_id = -1
else:
ipc_reply.app_id, ipc_reply.stage_id = self.env.spark_inverse_node_map[
node]
if node.idx not in node.job_dag.frontier_nodes:
# move (or stay) the executor to the job only
ipc_reply.stage_id = -1
if ipc_msg.app_id != 'void' and \
ipc_reply.app_id != 'void' and \
ipc_msg.app_id != ipc_reply.app_id:
# executor needs to move to another job, keep track of it
self.exec_tracker.add_executor_flow(ipc_reply.app_id, 1)
ipc_reply.msg = \
"external scheduler return app_id " + str(ipc_reply.app_id) + \
" stage_id " + str(ipc_reply.stage_id) + \
" for exec_id " + str(ipc_msg.exec_id)
elif ipc_msg.msg_type == 'deregister':
job_dag = self.env.remove_job_dag(ipc_msg.app_id)
remove_job_from_graph(self.graph, job_dag)
self.dag_db.remove_app(ipc_msg.app_id)
self.exec_tracker.remove_app(ipc_msg.app_id)
ipc_reply.msg = \
"external scheduler deregister app " + ipc_msg.app_id
print("time:", datetime.now())
print("msg_type:", ipc_msg.msg_type)
print("app_name:", ipc_msg.app_name)
print("app_id:", ipc_msg.app_id)
print("stage_id:", ipc_msg.stage_id)
print("executor_id:", ipc_msg.exec_id)
print("track_id:", ipc_msg.track_id)
print("num_available_executors:", ipc_msg.num_available_executors)
print("num_tasks_left", ipc_msg.num_tasks_left)
print("reply_msg:", ipc_reply.msg)
print("")
sys.stdout.flush()
socket.send(ipc_reply.SerializeToString())
class JobDAG(object):
def __init__(self, nodes, adj_mat, name):
# nodes: list of N nodes
# adj_mat: N by N 0-1 adjacency matrix, e_ij = 1 -> edge from i to j
assert len(nodes) == adj_mat.shape[0]
assert adj_mat.shape[0] == adj_mat.shape[1]
self.name = name
self.nodes = nodes
self.adj_mat = adj_mat
self.num_nodes = len(self.nodes)
self.num_nodes_done = 0
# set of executors currently running on the job
self.executors = OrderedSet()
# the computation graph needs to be a DAG
assert is_dag(self.num_nodes, self.adj_mat)
# get the set of schedule nodes
self.frontier_nodes = OrderedSet()
for node in self.nodes:
if node.is_schedulable():
self.frontier_nodes.add(node)
# assign job dag to node
self.assign_job_dag_to_node()
# dag is arrived
self.arrived = False
# dag is completed
self.completed = False
# dag start ime
self.start_time = None
# dag completion time
self.completion_time = np.inf
# map a executor number to an interval
self.executor_interval_map = \
self.get_executor_interval_map()
def assign_job_dag_to_node(self):
for node in self.nodes:
node.job_dag = self
def get_executor_interval_map(self):
executor_interval_map = {}
executor_data_point = [5, 10, 20, 40, 50, 60, 80, 100]
entry_pt = 0
# get the left most map
for e in range(executor_data_point[0] + 1):
executor_interval_map[e] = \
(executor_data_point[0],
executor_data_point[0])
# get the center map
for i in range(len(executor_data_point) - 1):
for e in range(executor_data_point[i] + 1,
executor_data_point[i + 1]):
executor_interval_map[e] = \
(executor_data_point[i],
executor_data_point[i + 1])
# at the data point
e = executor_data_point[i + 1]
executor_interval_map[e] = \
(executor_data_point[i + 1],
executor_data_point[i + 1])
# get the residual map
if config.exec_cap > executor_data_point[-1]:
for e in range(executor_data_point[-1] + 1, config.exec_cap + 1):
executor_interval_map[e] = \
(executor_data_point[-1],
executor_data_point[-1])
return executor_interval_map
def get_nodes_duration(self):
# Warning: this is slow O(num_nodes * num_tasks)
# get the duration over all nodes
duration = 0
for node in self.nodes:
duration += node.get_node_duration()
return duration
def reset(self):
for node in self.nodes:
node.reset()
self.num_nodes_done = 0
self.executors = OrderedSet()
self.frontier_nodes = OrderedSet()
for node in self.nodes:
if node.is_schedulable():
self.frontier_nodes.add(node)
self.arrived = False
self.completed = False
self.completion_time = np.inf
def update_frontier_nodes(self, node):
frontier_nodes_changed = False
for child in node.child_nodes:
if child.is_schedulable():
if child.idx not in self.frontier_nodes:
self.frontier_nodes.add(child)
frontier_nodes_changed = True
return frontier_nodes_changed
class Node(object):
def __init__(self, idx, tasks, task_duration, wall_time, np_random):
self.idx = idx
self.tasks = tasks
self.wall_time = wall_time
self.np_random = np_random
self.task_duration = task_duration
self.num_tasks = len(tasks)
self.num_finished_tasks = 0
self.next_task_idx = 0
self.no_more_tasks = False
self.tasks_all_done = False
self.node_finish_time = np.inf
self.executors = OrderedSet()
# uninitialized
self.parent_nodes = []
self.child_nodes = []
self.descendant_nodes = []
self.job_dag = None
self.assign_node_to_tasks()
def assign_node_to_tasks(self):
for task in self.tasks:
task.node = self
def get_node_duration(self):
# Warning: this is slow O(num_tasks)
# get the total duration over all tasks
duration = 0
for task in self.tasks:
duration += task.get_duration()
return duration
def is_schedulable(self):
if self.no_more_tasks: # no more tasks
return False
if self.tasks_all_done: # node done
return False
for node in self.parent_nodes:
if not node.tasks_all_done: # a parent node not done
return False
return True
def reset(self):
for task in self.tasks:
task.reset()
self.executors.clear()
self.num_finished_tasks = 0
self.next_task_idx = 0
self.no_more_tasks = False
self.tasks_all_done = False
self.node_finish_time = np.inf
def sample_executor_key(self, num_executors):
(left_exec, right_exec) = \
self.job_dag.executor_interval_map[num_executors]
executor_key = None
if left_exec == right_exec:
executor_key = left_exec
else:
rand_pt = self.np_random.randint(1, right_exec - left_exec + 1)
if rand_pt <= num_executors - left_exec:
executor_key = left_exec
else:
executor_key = right_exec
if executor_key not in self.task_duration['first_wave']:
# more executors than number of tasks in the job
largest_key = 0
for e in self.task_duration['first_wave']:
if e > largest_key:
largest_key = e
executor_key = largest_key
return executor_key
def schedule(self, executor):
assert self.next_task_idx < self.num_tasks
task = self.tasks[self.next_task_idx]
# task duration is determined by wave
num_executors = len(self.job_dag.executors)
assert num_executors > 0
# sample an executor point in the data
executor_key = self.sample_executor_key(num_executors)
if executor.task is None or \
executor.task.node.job_dag != task.node.job_dag:
# the executor never runs a task in this job
# fresh executor incurrs a warmup delay
if len(self.task_duration['fresh_durations'][executor_key]) > 0:
# (1) try to directly retrieve the warmup delay from data
fresh_durations = \
self.task_duration['fresh_durations'][executor_key]
i = np.random.randint(len(fresh_durations))
duration = fresh_durations[i]
else:
# (2) use first wave but deliberately add in a warmup delay
first_wave = \
self.task_duration['first_wave'][executor_key]
i = np.random.randint(len(first_wave))
duration = first_wave[i] + config.warmup_delay
elif executor.task is not None and \
executor.task.node == task.node and \
len(self.task_duration['rest_wave'][executor_key]) > 0:
# executor was working on this node
# the task duration should be retrieved from rest wave
rest_wave = self.task_duration['rest_wave'][executor_key]
i = np.random.randint(len(rest_wave))
duration = rest_wave[i]
else:
# executor is fresh to this node, use first wave
if len(self.task_duration['first_wave'][executor_key]) > 0:
# (1) try to retrieve first wave from data
first_wave = \
self.task_duration['first_wave'][executor_key]
i = np.random.randint(len(first_wave))
duration = first_wave[i]
else:
# (2) first wave doesn't exist, use fresh durations instead
# (should happen very rarely)
fresh_durations = \
self.task_duration['fresh_durations'][executor_key]
i = np.random.randint(len(fresh_durations))
duration = fresh_durations[i]
# # Hack! only use first/fresh duration
# # executor is fresh to this node, use first wave
# if len(self.task_duration['first_wave'][executor_key]) > 0:
# # (1) try to retrieve first wave from data
# first_wave = \
# self.task_duration['first_wave'][executor_key]
# i = self.np_random.randint(len(first_wave))
# duration = first_wave[i]
# else:
# # (2) first wave doesn't exist, use fresh durations instead
# # (should happen very rarely)
# fresh_durations = \
# self.task_duration['fresh_durations'][executor_key]
# i = self.np_random.randint(len(fresh_durations))
# duration = fresh_durations[i]
# detach the executor from old node
# the executor can run task means it is local
# to the job at this point
executor.detach_node()
# schedule the task
task.schedule(self.wall_time.curr_time, duration, executor)
# mark executor as running in the node
self.executors.add(executor)
executor.node = self
self.next_task_idx += 1
self.no_more_tasks = (self.next_task_idx >= self.num_tasks)
if self.no_more_tasks:
if self in self.job_dag.frontier_nodes:
self.job_dag.frontier_nodes.remove(self)
return task
def add_job(self, job):
self.free_executors[job] = OrderedSet()
def step(self, action):
assert self.action_space.contains(action)
next_node, limit_idx = action
# index starts from 0 but degree of parallelism starts with 1
limit = limit_idx + 1
# mark the node as selected
assert next_node not in self.node_selected
self.node_selected.add(next_node)
# commit the source executor
executor = next(iter(self.exec_to_schedule))
source = executor.job_dag if executor.node is None else executor.node
# compute number of valid executors to assign
if next_node is not None:
use_exec = min(next_node.num_tasks - next_node.next_task_idx - \
self.exec_commit.node_commit[next_node] - \
self.moving_executors.count(next_node), limit,
self.num_source_exec)
else:
use_exec = self.num_source_exec
assert use_exec > 0
self.exec_commit.add(source, next_node, use_exec)
# deduct the executors that know the destination
self.num_source_exec -= use_exec
assert self.num_source_exec >= 0
if self.num_source_exec == 0:
# now a new scheduling round, clean up node selection
self.node_selected.clear()
# all commitments are made, now schedule free executors
self.schedule()
# Now run to the next event in the virtual timeline
while len(self.timeline) > 0 and self.num_source_exec == 0:
# consult agent by putting executors in source_exec
new_time, obj = self.timeline.pop()
self.wall_time.update_time(new_time)
# case task: a task completion event, and frees up an executor.
# case query: a new job arrives
# case executor: an executor arrives at certain job
if isinstance(obj, Task): # task completion event
finished_task = obj
node = finished_task.node
node.num_finished_tasks += 1
# bookkeepings for node completion
frontier_changed = False
if node.num_finished_tasks == node.num_tasks:
assert not node.tasks_all_done # only complete once
node.tasks_all_done = True
node.job_dag.num_nodes_done += 1
node.node_finish_time = self.wall_time.curr_time
frontier_changed = node.job_dag.update_frontier_nodes(node)
# assign new destination for the job
self.assign_executor(finished_task.executor, frontier_changed)
# bookkeepings for job completion
if node.job_dag.num_nodes_done == node.job_dag.num_nodes:
assert not node.job_dag.completed # only complete once
node.job_dag.completed = True
node.job_dag.completion_time = self.wall_time.curr_time
self.remove_job(node.job_dag)
elif isinstance(obj, JobDAG): # new job arrival event
job_dag = obj
# job should be arrived at the first time
assert not job_dag.arrived
job_dag.arrived = True
# inform agent about job arrival when stream is enabled
self.job_dags.add(job_dag)
self.add_job(job_dag)
self.action_map = compute_act_map(self.job_dags)
# assign free executors (if any) to the new job
if len(self.free_executors[None]) > 0:
self.exec_to_schedule = \
OrderedSet(self.free_executors[None])
self.source_job = None
self.num_source_exec = \
len(self.free_executors[None])
elif isinstance(obj, Executor): # executor arrival event
executor = obj
# pop destination from the tracking record
node = self.moving_executors.pop(executor)
if node is not None:
# the job is not yet done when executor arrives
executor.job_dag = node.job_dag
node.job_dag.executors.add(executor)
if node is not None and not node.no_more_tasks:
# the node is still schedulable
if node in node.job_dag.frontier_nodes:
# node is immediately runnable
task = node.schedule(executor)
self.timeline.push(task.finish_time, task)
else:
# free up the executor in this job
self.free_executors.add(executor.job_dag, executor)
else:
# the node is saturated or the job is done
# by the time the executor arrives, use
# backup logic
self.backup_schedule(executor)
else:
print("illegal event type")
exit(1)
# compute reward
reward = self.reward_calculator.get_reward(
self.job_dags, self.wall_time.curr_time)
# no more decision to make, jobs all done or time is up
done = (self.num_source_exec == 0) and \
((len(self.timeline) == 0) or \
(self.wall_time.curr_time >= self.max_time))
if done:
assert self.wall_time.curr_time >= self.max_time or \
len(self.job_dags) == 0
return self.observe(), reward, done, None
class SparkSimEnv(core.Env):
"""
A trace-driven simulator for the dynamics of the scheduling module in Apache Spark.
The intricacies to closely simulate the system in reality are mainly (1) the "moving
cost" of executors acorss jobs (due to the overhead of starting a new JVM); (2) the
wave effect of running tasks of the same stage on an executor (overhead of loading
data in the first wave of tasks); (3) the dimenishing speedup in job runtime when
assigning more executors to a job. See reference for more details.
* STATE *
Graph type of observation. It consists of features associated with each node (
a tensor of dimension n * m, where n is number of nodes, m is number of features),
and adjacency matrix (a sparse 0-1 matrix of dimension n * n).
The features on each node is
[number_of_executors_currently_in_this_job, is_current_executor_local_to_this_job,
number_of_free_executors, total_work_remaining_on_this_node,
number_of_tasks_remaining_on_this_node]
* ACTIONS *
Two dimensional action, [node_idx_to_schedule_next, number_of_executors_to_assign]
Note: the set of available nodes has to contain node_idx, and the number of
executors to assign must not exceed the limit. Both the available set and the limit
are provided in the (auxiliary) state.
* REWARD *
Negative time elapsed for each job in the system since last action.
For example, the virtual time was 0 for the last action, 4 jobs
was in the system (either in the queue waiting or being processed),
job 1 finished at time 1, job 2 finished at time 2.4 and job 3 and 4
are still running at the next action. The next action is taken at
time 5. Then the reward is - (1 * 1 + 1 * 2.4 + 2 * 5).
Thus, the sum of the rewards would be negative of total
(waiting + processing) time for all jobs.
* REFERENCE *
Section 6.2
Learning Scheduling Algorithms for Data Processing Clusters
H Mao, M Schwarzkopf, SB Venkatakrishnan, M Alizadeh
https://arxiv.org/pdf/1810.01963.pdf
"""
def __init__(self):
# observation and action space
self.setup_space()
# random seed
self.seed(config.seed)
# global timer
self.wall_time = WallTime()
# uses priority queue
self.timeline = Timeline()
# executors
self.executors = OrderedSet()
for exec_id in range(config.exec_cap):
self.executors.add(Executor(exec_id))
# free executors
self.free_executors = FreeExecutors(self.executors)
# moving executors
self.moving_executors = MovingExecutors()
# executor commit
self.exec_commit = ExecutorCommit()
# prevent agent keeps selecting the same node
self.node_selected = set()
# for computing reward at each step
self.reward_calculator = RewardCalculator()
def add_job(self, job_dag):
self.moving_executors.add_job(job_dag)
self.free_executors.add_job(job_dag)
self.exec_commit.add_job(job_dag)
add_job_in_graph(self.graph, job_dag)
def assign_executor(self, executor, frontier_changed):
if executor.node is not None and not executor.node.no_more_tasks:
# keep working on the previous node
task = executor.node.schedule(executor)
self.timeline.push(task.finish_time, task)
else:
# need to move on to other nodes
if frontier_changed:
# frontier changed, need to consult all free executors
# note: executor.job_dag might change after self.schedule()
source_job = executor.job_dag
if len(self.exec_commit[executor.node]) > 0:
# directly fulfill the commitment
self.exec_to_schedule = {executor}
self.schedule()
else:
# free up the executor
self.free_executors.add(source_job, executor)
# then consult all free executors
self.exec_to_schedule = OrderedSet(self.free_executors[source_job])
self.source_job = source_job
self.num_source_exec = len(self.free_executors[source_job])
else:
# just need to schedule one current executor
self.exec_to_schedule = {executor}
# only care about executors on the node
if len(self.exec_commit[executor.node]) > 0:
# directly fulfill the commitment
self.schedule()
else:
# need to consult for ALL executors on the node
# Note: self.exec_to_schedule is immediate
# self.num_source_exec is for commit
# so len(self.exec_to_schedule) !=
# self.num_source_exec can happen
self.source_job = executor.job_dag
self.num_source_exec = len(executor.node.executors)
def backup_schedule(self, executor):
backup_scheduled = False
if executor.job_dag is not None:
# first try to schedule on current job
for node in executor.job_dag.frontier_nodes:
if not self.saturated(node):
# greedily schedule a frontier node
task = node.schedule(executor)
self.timeline.push(task.finish_time, task)
backup_scheduled = True
break
# then try to schedule on any available node
if not backup_scheduled:
schedulable_nodes = self.get_frontier_nodes()
if len(schedulable_nodes) > 0:
node = next(iter(schedulable_nodes))
self.timeline.push(
self.wall_time.curr_time + config.moving_delay, executor)
# keep track of moving executors
self.moving_executors.add(executor, node)
backup_scheduled = True
# at this point if nothing available, leave executor idle
if not backup_scheduled:
self.free_executors.add(executor.job_dag, executor)
def get_frontier_nodes(self):
# frontier nodes := unsaturated nodes with all parent nodes saturated
frontier_nodes = OrderedSet()
for job_dag in self.job_dags:
for node in job_dag.nodes:
if not node in self.node_selected and not self.saturated(node):
parents_saturated = True
for parent_node in node.parent_nodes:
if not self.saturated(parent_node):
parents_saturated = False
break
if parents_saturated:
frontier_nodes.add(node)
return frontier_nodes
def get_executor_limits(self):
# "minimum executor limit" for each job
# executor limit := {job_dag -> int}
executor_limit = {}
for job_dag in self.job_dags:
if self.source_job == job_dag:
curr_exec = self.num_source_exec
else:
curr_exec = 0
# note: this does not count in the commit and moving executors
executor_limit[job_dag] = len(job_dag.executors) - curr_exec
return executor_limit
def observe(self):
# valid set of nodes
frontier_nodes = self.get_frontier_nodes()
# sort out the exec_map (where the executors are)
exec_map = {}
for job_dag in self.job_dags:
exec_map[job_dag] = len(job_dag.executors)
# count in moving executors
for node in self.moving_executors.moving_executors.values():
exec_map[node.job_dag] += 1
# count in executor commit
for s in self.exec_commit.commit:
if isinstance(s, JobDAG):
j = s
elif isinstance(s, Node):
j = s.job_dag
elif s is None:
j = None
else:
print('source', s, 'unknown')
exit(1)
for n in self.exec_commit.commit[s]:
if n is not None and n.job_dag != j:
exec_map[n.job_dag] += self.exec_commit.commit[s][n]
for job_dag in self.job_dags:
for node in job_dag.nodes:
feature = np.zeros([6])
# number of executors already in the job
feature[0] = exec_map[job_dag]
# source executor is from the current job (locality)
feature[1] = job_dag is self.source_job
# number of source executors
feature[2] = self.num_source_exec
# remaining number of tasks in the node
feature[3] = node.num_tasks - node.next_task_idx
# average task duration of the node
feature[4] = node.tasks[-1].duration
# is the current node valid
feature[5] = node in frontier_nodes
# update feature in observation
self.graph.update_nodes({node: feature})
# update mask in the action space
self.action_space[0].update_valid_set(frontier_nodes)
# return the graph as observation
obs = self.graph
assert self.observation_space.contains(obs)
return obs
def saturated(self, node):
# frontier nodes := unsaturated nodes with all parent nodes saturated
anticipated_task_idx = node.next_task_idx + \
self.exec_commit.node_commit[node] + \
self.moving_executors.count(node)
# note: anticipated_task_idx can be larger than node.num_tasks
# when the tasks finish very fast before commitments are fulfilled
return anticipated_task_idx >= node.num_tasks
def schedule(self):
executor = next(iter(self.exec_to_schedule))
source = executor.job_dag if executor.node is None else executor.node
# schedule executors from the source until the commitment is fulfilled
while len(self.exec_commit[source]) > 0 and \
len(self.exec_to_schedule) > 0:
# keep fulfilling the commitment using free executors
node = self.exec_commit.pop(source)
executor = self.exec_to_schedule.pop()
# mark executor as in use if it was free executor previously
if self.free_executors.contain_executor(executor.job_dag, executor):
self.free_executors.remove(executor)
if node is None:
# the next node is explicitly silent, make executor ilde
if executor.job_dag is not None and \
any([not n.no_more_tasks for n in \
executor.job_dag.nodes]):
# mark executor as idle in its original job
self.free_executors.add(executor.job_dag, executor)
else:
# no where to assign, put executor in null pool
self.free_executors.add(None, executor)
elif not node.no_more_tasks:
# node is not currently saturated
if executor.job_dag == node.job_dag:
# executor local to the job
if node in node.job_dag.frontier_nodes:
# node is immediately runnable
task = node.schedule(executor)
self.timeline.push(task.finish_time, task)
else:
# put executor back in the free pool
self.free_executors.add(executor.job_dag, executor)
else:
# need to move executor
self.timeline.push(
self.wall_time.curr_time + config.moving_delay, executor)
# keep track of moving executors
self.moving_executors.add(executor, node)
else:
# node is already saturated, use backup logic
self.backup_schedule(executor)
def step(self, action):
assert self.action_space.contains(action)
next_node, limit_idx = action
# index starts from 0 but degree of parallelism starts with 1
limit = limit_idx + 1
# mark the node as selected
assert next_node not in self.node_selected
self.node_selected.add(next_node)
# commit the source executor
executor = next(iter(self.exec_to_schedule))
source = executor.job_dag if executor.node is None else executor.node
# compute number of valid executors to assign
if next_node is not None:
use_exec = min(next_node.num_tasks - next_node.next_task_idx - \
self.exec_commit.node_commit[next_node] - \
self.moving_executors.count(next_node), limit,
self.num_source_exec)
else:
use_exec = self.num_source_exec
assert use_exec > 0
self.exec_commit.add(source, next_node, use_exec)
# deduct the executors that know the destination
self.num_source_exec -= use_exec
assert self.num_source_exec >= 0
if self.num_source_exec == 0:
# now a new scheduling round, clean up node selection
self.node_selected.clear()
# all commitments are made, now schedule free executors
self.schedule()
# Now run to the next event in the virtual timeline
while len(self.timeline) > 0 and self.num_source_exec == 0:
# consult agent by putting executors in source_exec
new_time, obj = self.timeline.pop()
self.wall_time.update_time(new_time)
# case task: a task completion event, and frees up an executor.
# case query: a new job arrives
# case executor: an executor arrives at certain job
if isinstance(obj, Task): # task completion event
finished_task = obj
node = finished_task.node
node.num_finished_tasks += 1
# bookkeepings for node completion
frontier_changed = False
if node.num_finished_tasks == node.num_tasks:
assert not node.tasks_all_done # only complete once
node.tasks_all_done = True
node.job_dag.num_nodes_done += 1
node.node_finish_time = self.wall_time.curr_time
frontier_changed = node.job_dag.update_frontier_nodes(node)
# assign new destination for the job
self.assign_executor(finished_task.executor, frontier_changed)
# bookkeepings for job completion
if node.job_dag.num_nodes_done == node.job_dag.num_nodes:
assert not node.job_dag.completed # only complete once
node.job_dag.completed = True
node.job_dag.completion_time = self.wall_time.curr_time
self.remove_job(node.job_dag)
elif isinstance(obj, JobDAG): # new job arrival event
job_dag = obj
# job should be arrived at the first time
assert not job_dag.arrived
job_dag.arrived = True
# inform agent about job arrival when stream is enabled
self.job_dags.add(job_dag)
self.add_job(job_dag)
self.action_map = compute_act_map(self.job_dags)
# assign free executors (if any) to the new job
if len(self.free_executors[None]) > 0:
self.exec_to_schedule = \
OrderedSet(self.free_executors[None])
self.source_job = None
self.num_source_exec = \
len(self.free_executors[None])
elif isinstance(obj, Executor): # executor arrival event
executor = obj
# pop destination from the tracking record
node = self.moving_executors.pop(executor)
if node is not None:
# the job is not yet done when executor arrives
executor.job_dag = node.job_dag
node.job_dag.executors.add(executor)
if node is not None and not node.no_more_tasks:
# the node is still schedulable
if node in node.job_dag.frontier_nodes:
# node is immediately runnable
task = node.schedule(executor)
self.timeline.push(task.finish_time, task)
else:
# free up the executor in this job
self.free_executors.add(executor.job_dag, executor)
else:
# the node is saturated or the job is done
# by the time the executor arrives, use
# backup logic
self.backup_schedule(executor)
else:
print("illegal event type")
exit(1)
# compute reward
reward = self.reward_calculator.get_reward(
self.job_dags, self.wall_time.curr_time)
# no more decision to make, jobs all done or time is up
done = (self.num_source_exec == 0) and \
((len(self.timeline) == 0) or \
(self.wall_time.curr_time >= self.max_time))
if done:
assert self.wall_time.curr_time >= self.max_time or \
len(self.job_dags) == 0
return self.observe(), reward, done, None
def remove_job(self, job_dag):
for executor in list(job_dag.executors):
executor.detach_job()
self.exec_commit.remove_job(job_dag)
self.free_executors.remove_job(job_dag)
self.moving_executors.remove_job(job_dag)
self.job_dags.remove(job_dag)
self.finished_job_dags.add(job_dag)
remove_job_from_graph(self.graph, job_dag)
self.action_map = compute_act_map(self.job_dags)
def reset(self, max_time=np.inf):
# reset observation and action space
self.setup_space()
self.max_time = max_time
self.wall_time.reset()
self.timeline.reset()
self.exec_commit.reset()
self.moving_executors.reset()
self.reward_calculator.reset()
self.finished_job_dags = OrderedSet()
self.node_selected.clear()
for executor in self.executors:
executor.reset()
self.free_executors.reset(self.executors)
# generate a set of new jobs
self.job_dags = generate_jobs(
self.np_random, self.timeline, self.wall_time)
# map action to dag_idx and node_idx
self.action_map = compute_act_map(self.job_dags)
# add initial set of jobs in the system
for job_dag in self.job_dags:
self.add_job(job_dag)
# put all executors as source executors initially
self.source_job = None
self.num_source_exec = len(self.executors)
self.exec_to_schedule = OrderedSet(self.executors)
return self.observe()
def seed(self, seed):
self.np_random = seeding.np_random(seed)
def setup_space(self):
# Set up the observation and action space
# The boundary of the space may change if the dynamics is changed
# a warning message will show up every time e.g., the observation falls
# out of the observation space
self.graph = DirectedGraph()
self.obs_node_low = np.array([0] * 6)
self.obs_node_high = np.array([config.exec_cap, 1, config.exec_cap, 1000, 100000, 1])
self.obs_edge_low = self.obs_edge_high = np.array([]) # features on nodes only
self.observation_space = spaces.Graph(
node_feature_space=spaces.MultiBox(
low=self.obs_node_low,
high=self.obs_node_high,
dtype=np.float32),
edge_feature_space=spaces.MultiBox(
low=self.obs_edge_low,
high=self.obs_edge_high,
dtype=np.float32))
self.action_space = spaces.Tuple(
(spaces.NodeInGraph(self.graph),
spaces.MaskedDiscrete(config.num_servers)))
class Environment(object):
def __init__(self, dag_db):
self.dag_db = dag_db
self.job_dags = OrderedSet()
self.action_map = {} # action index -> node
self.available_executors = {}
self.last_trigger = None
# executors
self.executors = {}
for exec_id in range(config.exec_cap):
self.executors[exec_id] = Executor(exec_id)
# dynamically bind {app_id -> job_dag}
self.spark_dag_map = {}
# dynamically bind {job_dag -> app_id}
self.spark_inverse_dag_map = {}
# dynamically bind {(app_id, stage_id) -> node}
self.spark_node_map = {}
# dynamically bind {node -> (app_id, stage_id)}
self.spark_inverse_node_map = {}
# dynamically bind {app_id -> {exec_id -> re-usable track_id}}
self.exec_id_track_id_map = {}
def add_job_dag(self, app_id):
job_dag = self.dag_db.apps_map[app_id]
job_dag.arrived = True
self.job_dags.add(job_dag)
# update map for job_dag
self.spark_dag_map[app_id] = job_dag
self.spark_inverse_dag_map[job_dag] = app_id
# update exec_id track_id bind map
self.exec_id_track_id_map[app_id] = {}
# update map for node
node_idx_to_stage_id_map = self.dag_db.stage_map[app_id]
for node in job_dag.nodes:
stage_id = node_idx_to_stage_id_map[node.idx]
self.spark_node_map[(app_id, stage_id)] = node
self.spark_inverse_node_map[node] = (app_id, stage_id)
# update map for actions
self.action_map.clear()
self.action_map.update(self.pre_compute_action_map())
return job_dag
def bind_exec_id(self, app_id, exec_id, track_id):
assert 0 <= track_id < config.exec_cap
self.exec_id_track_id_map[app_id][exec_id] = track_id
def complete_stage(self, app_id, stage_id):
node = self.spark_node_map[(app_id, stage_id)]
# bookkeepings for node completion
assert not node.tasks_all_done # only complete once
node.tasks_all_done = True
node.job_dag.update_frontier_nodes(node)
node.job_dag.num_nodes_done += 1
# bookkeepings for job completion
if node.job_dag.num_nodes_done == node.job_dag.num_nodes:
assert not node.job_dag.completed # only complete once
node.job_dag.completed = True
def complete_tasks(self, app_id, stage_id, num_tasks_left):
node = self.spark_node_map[(app_id, stage_id)]
prev_finished_tasks = node.num_finished_tasks
# update number of finished tasks for the node
node.num_finished_tasks = node.num_tasks - num_tasks_left
# update the next task index of the node
node.next_task_idx += node.num_finished_tasks - prev_finished_tasks
# remove node from frontier node if it is saturated
node.no_more_tasks = (node.next_task_idx >= node.num_tasks)
if node.no_more_tasks:
if node.idx in node.job_dag.frontier_nodes:
del node.job_dag.frontier_nodes[node.idx]
def pre_compute_action_map(self):
# translate action ~ [0, num_nodes_in_all_dags) to node object
action_map = {}
action = 0
for job_dag in self.job_dags:
for node in job_dag.nodes:
action_map[action] = node
action += 1
return action_map
def remove_job_dag(self, app_id):
job_dag = self.dag_db.apps_map[app_id]
self.job_dags.remove(job_dag)
# free up stage holding executors
for executor in job_dag.executors:
executor.task = None
executor.job_dag = None
# update exec_id track_id map
del self.exec_id_track_id_map[app_id]
# update map for job_dag
del self.spark_dag_map[app_id]
del self.spark_inverse_dag_map[job_dag]
# update map for node
node_idx_to_stage_id_map = self.dag_db.stage_map[app_id]
for node in job_dag.nodes:
stage_id = node_idx_to_stage_id_map[node.idx]
del self.spark_node_map[(app_id, stage_id)]
del self.spark_inverse_node_map[node]
# update map for actions
self.action_map.clear()
self.action_map.update(self.pre_compute_action_map())
return job_dag
def reset(self, executors):
self.free_executors = {}
self.free_executors[None] = OrderedSet()
for executor in executors:
self.free_executors[None].add(executor)