class ReverseEasyScheduler(EasyBackfillScheduler):
def __init__(self, options):
super(ReverseEasyScheduler, self).__init__(options)
self.cpu_snapshot = CpuSnapshot(self.num_processors, options["stats"])
def _backfill_jobs(self, current_time):
"Overriding parent method"
if len(self.unscheduled_jobs) <= 1:
return []
result = []
first_job = self.unscheduled_jobs[0]
tail = list_copy(self.unscheduled_jobs[1:])
tail_of_jobs_by_reverse_order = sorted(tail, key=latest_sort_key)
self.cpu_snapshot.assignJobEarliest(first_job, current_time)
for job in tail_of_jobs_by_reverse_order:
if self.cpu_snapshot.canJobStartNow(job, current_time):
self.unscheduled_jobs.remove(job)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
self.cpu_snapshot.delJobFromCpuSlices(first_job)
return result
class TailDoubleEasyScheduler(EasyBackfillScheduler):
""" This algorithm implements the algorithm in the paper of Tsafrir, Etzion, Feitelson, june 2007?
"""
def __init__(self, options):
super(TailDoubleEasyScheduler, self).__init__(options)
self.cpu_snapshot = CpuSnapshot(self.num_processors, options["stats"])
def _backfill_jobs(self, current_time):
"Overriding parent method"
if len(self.unscheduled_jobs) <= 1:
return []
result = []
first_job = self.unscheduled_jobs[0]
tail = list_copy(self.unscheduled_jobs[1:])
self.cpu_snapshot.assignJobEarliest(first_job, current_time)
for job in tail:
job.predicted_run_time = 2 * job.user_estimated_run_time # doubling is done here
if self.cpu_snapshot.canJobStartNow(job, current_time): # if job can be backfilled
self.unscheduled_jobs.remove(job)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
else:
job.predicted_run_time = job.user_estimated_run_time # undoubling is done here
self.cpu_snapshot.delJobFromCpuSlices(first_job)
return result
class ConservativeScheduler(Scheduler):
def __init__(self, options):
super(ConservativeScheduler, self).__init__(options)
self.cpu_snapshot = CpuSnapshot(self.num_processors, options["stats"])
self.unfinished_jobs_by_submit_time = []
def new_events_on_job_submission(self, job, current_time):
self.cpu_snapshot.archive_old_slices(current_time)
self.unfinished_jobs_by_submit_time.append(job)
self.cpu_snapshot.assignJobEarliest(job, current_time)
return [JobStartEvent(job.start_to_run_at_time, job)]
def new_events_on_job_termination(self, job, current_time):
""" Here we delete the tail of job if it was ended before the duration declaration.
It then reschedules the remaining jobs and returns a collection of new termination events
(using the dictionary data structure) """
self.cpu_snapshot.archive_old_slices(current_time)
self.unfinished_jobs_by_submit_time.remove(job)
self.cpu_snapshot.delTailofJobFromCpuSlices(job)
return self._reschedule_jobs(current_time)
def _reschedule_jobs(self, current_time):
newEvents = []
for job in self.unfinished_jobs_by_submit_time:
if job.start_to_run_at_time <= current_time:
continue # job started to run before, so it cannot be rescheduled (preemptions are not allowed)
prev_start_to_run_at_time = job.start_to_run_at_time
self.cpu_snapshot.delJobFromCpuSlices(job)
self.cpu_snapshot.assignJobEarliest(job, current_time)
assert prev_start_to_run_at_time >= job.start_to_run_at_time
if prev_start_to_run_at_time != job.start_to_run_at_time:
newEvents.append(JobStartEvent(job.start_to_run_at_time, job))
return newEvents
class ConservativeScheduler(Scheduler):
def __init__(self, options):
super(ConservativeScheduler, self).__init__(options)
self.cpu_snapshot = CpuSnapshot(self.num_processors, options["stats"])
self.unfinished_jobs_by_submit_time = []
def new_events_on_job_submission(self, job, current_time):
self.cpu_snapshot.archive_old_slices(current_time)
self.unfinished_jobs_by_submit_time.append(job)
self.cpu_snapshot.assignJobEarliest(job, current_time)
return [ JobStartEvent(job.start_to_run_at_time, job) ]
def new_events_on_job_termination(self, job, current_time):
""" Here we delete the tail of job if it was ended before the duration declaration.
It then reschedules the remaining jobs and returns a collection of new termination events
(using the dictionary data structure) """
self.cpu_snapshot.archive_old_slices(current_time)
self.unfinished_jobs_by_submit_time.remove(job)
self.cpu_snapshot.delTailofJobFromCpuSlices(job)
return self._reschedule_jobs(current_time)
def _reschedule_jobs(self, current_time):
newEvents = []
for job in self.unfinished_jobs_by_submit_time:
if job.start_to_run_at_time <= current_time:
continue # job started to run before, so it cannot be rescheduled (preemptions are not allowed)
prev_start_to_run_at_time = job.start_to_run_at_time
self.cpu_snapshot.delJobFromCpuSlices(job)
self.cpu_snapshot.assignJobEarliest(job, current_time)
assert prev_start_to_run_at_time >= job.start_to_run_at_time
if prev_start_to_run_at_time != job.start_to_run_at_time:
newEvents.append( JobStartEvent(job.start_to_run_at_time, job) )
return newEvents
class EasySjbfScheduler(EasyBackfillScheduler):
""" This algorithm implements the algorithm in the paper of Tsafrir, Etzion, Feitelson, june 2007?
"""
def __init__(self, options):
super(EasySjbfScheduler, self).__init__(options)
self.cpu_snapshot = CpuSnapshot(self.num_processors, options["stats"])
def _backfill_jobs(self, current_time):
"Overriding parent method"
if len(self.unscheduled_jobs) <= 1:
return []
result = []
first_job = self.unscheduled_jobs[0]
tail = list_copy(self.unscheduled_jobs[1:])
tail_of_jobs_by_sjf_order = sorted(tail, key=sjf_sort_key)
self.cpu_snapshot.assignJobEarliest(first_job, current_time)
for job in tail_of_jobs_by_sjf_order:
if self.cpu_snapshot.canJobStartNow(job, current_time):
job.is_backfilled = 1
self.unscheduled_jobs.remove(job)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
self.cpu_snapshot.unAssignJob(first_job)
return result
class EasySJBFScheduler(EasyBackfillScheduler):
""" This algorithm implements the algorithm in the paper of Tsafrir, Etzion, Feitelson, june 2007?
"""
def __init__(self, num_processors):
super(EasySJBFScheduler, self).__init__(num_processors)
self.cpu_snapshot = CpuSnapshot(num_processors)
def _backfill_jobs(self, current_time):
"Overriding parent method"
if len(self.unscheduled_jobs) <= 1:
return []
result = []
first_job = self.unscheduled_jobs[0]
tail = list_copy(self.unscheduled_jobs[1:])
tail_of_jobs_by_sjf_order = sorted(tail, key=sjf_sort_key)
self.cpu_snapshot.assignJobEarliest(first_job, current_time)
for job in tail_of_jobs_by_sjf_order:
if self.cpu_snapshot.canJobStartNow(job, current_time):
self.unscheduled_jobs.remove(job)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
self.cpu_snapshot.delJobFromCpuSlices(first_job)
return result
class EasyBackfillScheduler(Scheduler):
def __init__(self, options):
super(EasyBackfillScheduler, self).__init__(options)
self.cpu_snapshot = CpuSnapshot(self.num_processors, options["stats"])
self.unscheduled_jobs = []
def new_events_on_job_submission(self, just_submitted_job, current_time):
""" Here we first add the new job to the waiting list. We then try to schedule
the jobs in the waiting list, returning a collection of new termination events """
# TODO: a probable performance bottleneck because we reschedule all the
# jobs. Knowing that only one new job is added allows more efficient
# scheduling here.
self.cpu_snapshot.archive_old_slices(current_time)
self.unscheduled_jobs.append(just_submitted_job)
retl = []
if (self.cpu_snapshot.free_processors_available_at(current_time) >= just_submitted_job.num_required_processors):
for job in self._schedule_jobs(current_time):
retl.append(JobStartEvent(current_time, job))
return retl
def new_events_on_job_termination(self, job, current_time):
""" Here we first delete the tail of the just terminated job (in case it's
done before user estimation time). We then try to schedule the jobs in the waiting list,
returning a collection of new termination events """
self.cpu_snapshot.archive_old_slices(current_time)
self.cpu_snapshot.delTailofJobFromCpuSlices(job)
return [
JobStartEvent(current_time, job)
for job in self._schedule_jobs(current_time)
]
def _schedule_jobs(self, current_time):
"Schedules jobs that can run right now, and returns them"
jobs = self._schedule_head_of_list(current_time)
jobs += self._backfill_jobs(current_time)
return jobs
def _schedule_head_of_list(self, current_time):
result = []
while True:
if len(self.unscheduled_jobs) == 0:
break
# Try to schedule the first job
if self.cpu_snapshot.free_processors_available_at(current_time) >= self.unscheduled_jobs[0].num_required_processors:
job = self.unscheduled_jobs.pop(0)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
else:
# first job can't be scheduled
break
return result
def _backfill_jobs(self, current_time):
"""
Find jobs that can be backfilled and update the cpu snapshot.
DEPRECATED FUNCTION !!!!!!
"""
if len(self.unscheduled_jobs) <= 1:
return []
result = []
tail_of_waiting_list = list_copy(self.unscheduled_jobs[1:])
first_job = self.unscheduled_jobs[0]
self.cpu_snapshot.assignJobEarliest(first_job, current_time)
for job in tail_of_waiting_list:
if self.cpu_snapshot.canJobStartNow(job, current_time):
job.is_backfilled = 1
self.unscheduled_jobs.remove(job)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
self.cpu_snapshot.unAssignJob(first_job)
return result
class EasyPlusPlusScheduler(Scheduler):
""" This algorithm implements the algorithm in the paper of Tsafrir, Etzion, Feitelson, june 2007?
"""
I_NEED_A_PREDICTOR = True
def __init__(self, options):
super(EasyPlusPlusScheduler, self).__init__(options)
self.init_predictor(options)
self.init_corrector(options)
self.cpu_snapshot = CpuSnapshot(self.num_processors, options["stats"])
self.unscheduled_jobs = []
def new_events_on_job_submission(self, job, current_time):
self.cpu_snapshot.archive_old_slices(current_time)
self.predictor.predict(job, current_time, self.running_jobs)
if not hasattr(job,"initial_prediction"):
job.initial_prediction=job.predicted_run_time
self.unscheduled_jobs.append(job)
return [
JobStartEvent(current_time, job)
for job in self._schedule_jobs(current_time)
]
def new_events_on_job_termination(self, job, current_time):
self.predictor.fit(job, current_time)
if self.corrector.__name__=="ninetynine":
self.pestimator.fit(job.actual_run_time/job.user_estimated_run_time)
self.cpu_snapshot.archive_old_slices(current_time)
self.cpu_snapshot.delTailofJobFromCpuSlices(job)
return [
JobStartEvent(current_time, job)
for job in self._schedule_jobs(current_time)
]
def new_events_on_job_under_prediction(self, job, current_time):
pass #assert job.predicted_run_time <= job.user_estimated_run_time
if not hasattr(job,"num_underpredict"):
job.num_underpredict = 0
else:
job.num_underpredict += 1
if self.corrector.__name__=="ninetynine":
new_predicted_run_time = self.corrector(self.pestimator,job,current_time)
else:
new_predicted_run_time = self.corrector(job, current_time)
#set the new predicted runtime
self.cpu_snapshot.assignTailofJobToTheCpuSlices(job, new_predicted_run_time)
job.predicted_run_time = new_predicted_run_time
return [JobStartEvent(current_time, job)]
def _schedule_jobs(self, current_time):
"Schedules jobs that can run right now, and returns them"
jobs = self._schedule_head_of_list(current_time)
jobs += self._backfill_jobs(current_time)
return jobs
def _schedule_head_of_list(self, current_time):
result = []
while True:
if len(self.unscheduled_jobs) == 0:
break
# Try to schedule the first job
if self.cpu_snapshot.free_processors_available_at(current_time) >= self.unscheduled_jobs[0].num_required_processors:
job = self.unscheduled_jobs.pop(0)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
else:
# first job can't be scheduled
break
return result
def _backfill_jobs(self, current_time):
if len(self.unscheduled_jobs) <= 1:
return []
result = []
first_job = self.unscheduled_jobs[0]
tail = list_copy(self.unscheduled_jobs[1:])
tail_of_jobs_by_sjf_order = sorted(tail, key=sjf_sort_key)
self.cpu_snapshot.assignJobEarliest(first_job, current_time)
for job in tail_of_jobs_by_sjf_order:
if self.cpu_snapshot.canJobStartNow(job, current_time):
job.is_backfilled = 1
self.unscheduled_jobs.remove(job)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
self.cpu_snapshot.delJobFromCpuSlices(first_job)
return result
class EasyPlusPlusScheduler(Scheduler):
""" This algorithm implements the algorithm in the paper of Tsafrir, Etzion, Feitelson, june 2007?
"""
def __init__(self, num_processors):
super(EasyPlusPlusScheduler, self).__init__(num_processors)
self.cpu_snapshot = CpuSnapshot(num_processors)
self.unscheduled_jobs = []
self.user_run_time_prev = {}
self.user_run_time_last = {}
def new_events_on_job_submission(self, job, current_time):
if not self.user_run_time_last.has_key(job.user_id):
self.user_run_time_prev[job.user_id] = None
self.user_run_time_last[job.user_id] = None
self.cpu_snapshot.archive_old_slices(current_time)
self.unscheduled_jobs.append(job)
return [
JobStartEvent(current_time, job)
for job in self._schedule_jobs(current_time)
]
def new_events_on_job_termination(self, job, current_time):
assert self.user_run_time_last.has_key(job.user_id) == True
assert self.user_run_time_prev.has_key(job.user_id) == True
self.user_run_time_prev[job.user_id] = self.user_run_time_last[
job.user_id]
self.user_run_time_last[job.user_id] = job.actual_run_time
self.cpu_snapshot.archive_old_slices(current_time)
self.cpu_snapshot.delTailofJobFromCpuSlices(job)
return [
JobStartEvent(current_time, job)
for job in self._schedule_jobs(current_time)
]
def new_events_on_job_under_prediction(self, job, current_time):
assert job.predicted_run_time <= job.user_estimated_run_time
self.cpu_snapshot.assignTailofJobToTheCpuSlices(job)
job.predicted_run_time = job.user_estimated_run_time
return []
def _schedule_jobs(self, current_time):
"Schedules jobs that can run right now, and returns them"
for job in self.unscheduled_jobs:
if self.user_run_time_prev[job.user_id] != None:
average = int((self.user_run_time_last[job.user_id] +
self.user_run_time_prev[job.user_id]) / 2)
job.predicted_run_time = min(job.user_estimated_run_time,
average)
jobs = self._schedule_head_of_list(current_time)
jobs += self._backfill_jobs(current_time)
return jobs
def _schedule_head_of_list(self, current_time):
result = []
while True:
if len(self.unscheduled_jobs) == 0:
break
# Try to schedule the first job
if self.cpu_snapshot.free_processors_available_at(
current_time
) >= self.unscheduled_jobs[0].num_required_processors:
job = self.unscheduled_jobs.pop(0)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
else:
# first job can't be scheduled
break
return result
def _backfill_jobs(self, current_time):
if len(self.unscheduled_jobs) <= 1:
return []
result = []
first_job = self.unscheduled_jobs[0]
tail = list_copy(self.unscheduled_jobs[1:])
tail_of_jobs_by_sjf_order = sorted(tail, key=sjf_sort_key)
self.cpu_snapshot.assignJobEarliest(first_job, current_time)
for job in tail_of_jobs_by_sjf_order:
if self.cpu_snapshot.canJobStartNow(job, current_time):
self.unscheduled_jobs.remove(job)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
self.cpu_snapshot.delJobFromCpuSlices(first_job)
return result
class EasyPlusPlusScheduler(Scheduler):
""" This algorithm implements the algorithm in the paper of Tsafrir, Etzion, Feitelson, june 2007?
"""
def __init__(self, num_processors):
super(EasyPlusPlusScheduler, self).__init__(num_processors)
self.cpu_snapshot = CpuSnapshot(num_processors)
self.unscheduled_jobs = []
self.user_run_time_prev = {}
self.user_run_time_last = {}
def new_events_on_job_submission(self, job, current_time):
if not self.user_run_time_last.has_key(job.user_id):
self.user_run_time_prev[job.user_id] = None
self.user_run_time_last[job.user_id] = None
self.cpu_snapshot.archive_old_slices(current_time)
self.unscheduled_jobs.append(job)
return [
JobStartEvent(current_time, job)
for job in self._schedule_jobs(current_time)
]
def new_events_on_job_termination(self, job, current_time):
assert self.user_run_time_last.has_key(job.user_id) == True
assert self.user_run_time_prev.has_key(job.user_id) == True
self.user_run_time_prev[job.user_id] = self.user_run_time_last[job.user_id]
self.user_run_time_last[job.user_id] = job.actual_run_time
self.cpu_snapshot.archive_old_slices(current_time)
self.cpu_snapshot.delTailofJobFromCpuSlices(job)
return [
JobStartEvent(current_time, job)
for job in self._schedule_jobs(current_time)
]
def new_events_on_job_under_prediction(self, job, current_time):
assert job.predicted_run_time <= job.user_estimated_run_time
self.cpu_snapshot.assignTailofJobToTheCpuSlices(job)
job.predicted_run_time = job.user_estimated_run_time
return []
def _schedule_jobs(self, current_time):
"Schedules jobs that can run right now, and returns them"
for job in self.unscheduled_jobs:
if self.user_run_time_prev[job.user_id] != None:
average = int((self.user_run_time_last[job.user_id] + self.user_run_time_prev[job.user_id])/ 2)
job.predicted_run_time = min (job.user_estimated_run_time, average)
jobs = self._schedule_head_of_list(current_time)
jobs += self._backfill_jobs(current_time)
return jobs
def _schedule_head_of_list(self, current_time):
result = []
while True:
if len(self.unscheduled_jobs) == 0:
break
# Try to schedule the first job
if self.cpu_snapshot.free_processors_available_at(current_time) >= self.unscheduled_jobs[0].num_required_processors:
job = self.unscheduled_jobs.pop(0)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
else:
# first job can't be scheduled
break
return result
def _backfill_jobs(self, current_time):
if len(self.unscheduled_jobs) <= 1:
return []
result = []
first_job = self.unscheduled_jobs[0]
tail = list_copy(self.unscheduled_jobs[1:])
tail_of_jobs_by_sjf_order = sorted(tail, key=sjf_sort_key)
self.cpu_snapshot.assignJobEarliest(first_job, current_time)
for job in tail_of_jobs_by_sjf_order:
if self.cpu_snapshot.canJobStartNow(job, current_time):
self.unscheduled_jobs.remove(job)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
self.cpu_snapshot.delJobFromCpuSlices(first_job)
return result
class EasyPlusPlusScheduler(Scheduler):
""" This algorithm implements the algorithm in the paper of Tsafrir, Etzion, Feitelson, june 2007?
"""
I_NEED_A_PREDICTOR = True
def __init__(self, options):
super(EasyPlusPlusScheduler, self).__init__(options)
self.init_predictor(options)
self.init_corrector(options)
self.cpu_snapshot = CpuSnapshot(self.num_processors, options["stats"])
self.unscheduled_jobs = []
self.ff = open("times-epp-sgd.txt", 'w')
def new_events_on_job_submission(self, job, current_time):
self.cpu_snapshot.archive_old_slices(current_time)
self.predictor.predict(job, current_time, self.running_jobs)
self.ff.write("%d\t%d\n" %
(job.actual_run_time, job.predicted_run_time))
self.ff.flush()
if not hasattr(job, "initial_prediction"):
job.initial_prediction = job.predicted_run_time
self.unscheduled_jobs.append(job)
return [
JobStartEvent(current_time, job)
for job in self._schedule_jobs(current_time)
]
def new_events_on_job_termination(self, job, current_time):
self.predictor.fit(job, current_time)
if self.corrector.__name__ == "ninetynine":
self.pestimator.fit(job.actual_run_time /
job.user_estimated_run_time)
self.cpu_snapshot.archive_old_slices(current_time)
self.cpu_snapshot.delTailofJobFromCpuSlices(job)
return [
JobStartEvent(current_time, job)
for job in self._schedule_jobs(current_time)
]
def new_events_on_job_under_prediction(self, job, current_time):
assert job.predicted_run_time <= job.user_estimated_run_time
if not hasattr(job, "num_underpredict"):
job.num_underpredict = 0
else:
job.num_underpredict += 1
if self.corrector.__name__ == "ninetynine":
new_predicted_run_time = self.corrector(self.pestimator, job,
current_time)
else:
new_predicted_run_time = self.corrector(job, current_time)
#set the new predicted runtime
self.cpu_snapshot.assignTailofJobToTheCpuSlices(
job, new_predicted_run_time)
job.predicted_run_time = new_predicted_run_time
return [JobStartEvent(current_time, job)]
def _schedule_jobs(self, current_time):
"Schedules jobs that can run right now, and returns them"
jobs = self._schedule_head_of_list(current_time)
jobs += self._backfill_jobs(current_time)
return jobs
def _schedule_head_of_list(self, current_time):
result = []
while True:
if len(self.unscheduled_jobs) == 0:
break
# Try to schedule the first job
if self.cpu_snapshot.free_processors_available_at(
current_time
) >= self.unscheduled_jobs[0].num_required_processors:
job = self.unscheduled_jobs.pop(0)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
else:
# first job can't be scheduled
break
return result
def _backfill_jobs(self, current_time):
if len(self.unscheduled_jobs) <= 1:
return []
result = []
first_job = self.unscheduled_jobs[0]
tail = list_copy(self.unscheduled_jobs[1:])
tail_of_jobs_by_sjf_order = sorted(tail, key=sjf_sort_key)
self.cpu_snapshot.assignJobEarliest(first_job, current_time)
for job in tail_of_jobs_by_sjf_order:
if self.cpu_snapshot.canJobStartNow(job, current_time):
job.is_backfilled = 1
self.unscheduled_jobs.remove(job)
self.cpu_snapshot.assignJob(job, current_time)
result.append(job)
self.cpu_snapshot.delJobFromCpuSlices(first_job)
return result
