def execute_query(aws_key_id, aws_key, args, query, branch, is_first_branch):
"""
Executes the specified query using the branch that is currently checked out (whose name is
`branch`). Copies the event log and continuous monitors file to `args.output_dir`.
`is_first_branch` should be set to True if this is the first branch to be tested.
"""
print_heading("Executing {} {} of query {} using branch '{}'".format(
args.num_trials, "trial" if (args.num_trials == 1) else "trials",
query, branch))
# Restart Spark and the Thrift server in order to make sure that we are using the correct version,
# and to start using new log files.
stop_thriftserver()
stop_spark()
start_spark()
benchmark_runner_dir = path.join(args.benchmark_dir, "runner")
driver_addr = subprocess.check_output(
"curl -s http://169.254.169.254/latest/meta-data/public-hostname",
shell=True)
print "Creating benchmark tables and starting the Thrift server"
prepare_benchmark_script = path.join(benchmark_runner_dir,
"prepare-benchmark.sh")
prepare_benchmark_command = "{} \
--spark \
--aws-key-id={} \
--aws-key={} \
--spark-host={} \
--spark-identity-file={} \
--scale-factor={} \
--file-format={} \
--skip-s3-import".format(prepare_benchmark_script, aws_key_id, aws_key,
driver_addr, args.identity_file,
args.scale_factor, args.file_format)
if args.parquet:
prepare_benchmark_command += " --parquet"
if not is_first_branch or args.skip_parquet_conversion:
prepare_benchmark_command += " --skip-parquet-conversion"
execute_shell_command(prepare_benchmark_command)
if args.memory:
cache_table_for_query(query)
print "Executing query"
run_query_script = path.join(benchmark_runner_dir, "run-query.sh")
run_query_command = "{} \
--spark \
--spark-host={} \
--spark-identity-file={} \
--query-num={} \
--num-trials={} \
--clear-buffer-cache".format(run_query_script, driver_addr,
args.identity_file, query, args.num_trials)
if args.compress_output:
run_query_command += " --compress"
if args.memory:
run_query_command += " --spark-cache-output-tables"
execute_shell_command(run_query_command)
# Stop the Thrift server and Spark in order to stop using the Spark log files.
stop_thriftserver()
stop_spark()
print "Retrieving logs"
parameters = [query, branch]
log_dir = utils.copy_all_logs(parameters, utils.get_workers())
utils.copy_all_traces(log_dir, driver_addr, utils.get_workers())
log_files = path.join(log_dir, "*")
# Move the logs into a new directory: output_dir/query/branch/
output_dir = path.join(args.output_dir, query, branch)
execute_shell_command("mkdir -pv {}".format(output_dir))
execute_shell_command("mv -v {} {}".format(log_files, output_dir))
execute_shell_command("rm -rf {}".format(log_dir))
def solver(data):
# initialise solver
solver = pywrapcp.Solver("allocations")
tasks = utils.get_tasks(data['scheduledTasks'])
workers = utils.get_workers(data['workers'])
cost_matrix = data['costMatrix']
solver_option = data['solverOption']
time_limit = data['timeLimit']
extra_constraints = data['constraints'] if 'constraints' in data else {}
print('solver_option', solver_option)
num_tasks = len(tasks)
num_workers = len(workers)
# declare decision variables and a reference matrix
assignment_costs = []
assignments = []
assignments_ref = []
for worker in workers:
worker_assignments = []
worker_assignments_ref = []
worker_assignment_costs = []
for task in tasks:
worker_assignments.append(
solver.IntVar(0, 1, f'worker: , task: {task.id}'))
worker_assignments_ref.append(Worker_task(worker, task))
worker_assignment_costs.append(cost_matrix[str(
worker.id)][task.id])
assignments.append(worker_assignments)
assignments_ref.append(worker_assignments_ref)
assignment_costs.append(worker_assignment_costs)
constraints = Constraints(
tasks,
workers,
assignment_costs,
assignments,
assignments_ref,
)
# objective
# Only add objective if optimisation requested
if solver_option != 'noOptimisation':
total_cost = solver.IntVar(0, 3000, "total_cost")
solver.Add(total_cost == solver.Sum([
assignment_costs[i][j] * assignments[i][j]
for i in range(num_workers) for j in range(num_tasks)
]))
objective = solver.Minimize(total_cost, 5)
# constraints
# each task assigned it's given qty
constraints.add_task_qty_constraint(solver)
# a worker cannot work on two tasks that are on at the same time
constraints.add_same_worker_same_task_time(solver)
# a worker can at most be assigned to the same orderTask date once (i.e cannot take up multiple qty)
# maybe add any cannot work constraints
# maybe add any must work constraints
must_map = extra_constraints[
'mustWork'] if 'mustWork' in extra_constraints else None
cannot_map = extra_constraints[
'cannotWork'] if 'cannotWork' in extra_constraints else None
constraints.must_cannot_work(solver, must_map, cannot_map)
# add must combined must work
if 'combinedMustWork' in extra_constraints:
constraints.combined_must_work_all(
solver, extra_constraints['combinedMustWork'])
# add at least has to work constraint
if 'atLeastWork' in extra_constraints:
constraints.add_at_least_work_task(solver,
extra_constraints['atLeastWork'])
# add total time fatigue constraints
if 'timeFatigueTotal' in extra_constraints:
constraints.add_time_fatigue_total(
solver, extra_constraints['timeFatigueTotal'])
# add total overall time fatigue constraints
if 'overallTimeFatigueTotal' in extra_constraints:
constraints.add_overall_total_fatigue_time(
solver, extra_constraints['overallTimeFatigueTotal'])
# add consecutive fatigue constaints
if 'overallTimeFatigueConsecutive' in extra_constraints:
constraints.add_overall_consecutive_total_fatigue_time(
solver, extra_constraints['overallTimeFatigueConsecutive'])
# add unavailable time constraints
if 'unavailable' in extra_constraints:
constraints.add_unavailability(solver,
extra_constraints['unavailable'])
# add buddy constraints
if 'buddy' in extra_constraints:
constraints.add_buddy(solver, extra_constraints['buddy'])
# add nemesis constraints
if 'nemesis' in extra_constraints:
constraints.add_nemesis(solver, extra_constraints['nemesis'])
# works must be assigned to at least n tasks (this could change later per worker)
# [solver.Add(solver.Sum(assignments[i][j] for j in range(num_tasks)) >= 3) for i in range(num_workers)]
# Create the decision builder.
# Want to sort the decision variables by least cost to the solution
if solver_option != 'noOptimisation':
assignment_ref_copy = copy.deepcopy(assignments_ref)
assignment_ref_copy_flat = [
assignment_ref_copy[i][j] for i in range(num_workers)
for j in range(num_tasks)
]
# Sort by least cost
assignment_ref_copy_flat.sort(key=lambda wrk_tsk: cost_matrix[str(
wrk_tsk.worker.id)][wrk_tsk.task.id])
# map to assignment vars
assignments_flat = [
assignments[ref.worker.index][ref.task.index]
for ref in assignment_ref_copy_flat
]
else:
assignments_flat = [
assignments[i][j] for i in range(num_workers)
for j in range(num_tasks)
]
db = solver.Phase(assignments_flat, solver.CHOOSE_FIRST_UNBOUND,
solver.ASSIGN_MAX_VALUE)
# Create solution collector depending on solver option requested
if (solver_option == 'optimise'
and time_limit != None) or solver_option == 'optimal':
collector = solver.BestValueSolutionCollector(
False) # False finds minimum as best solution
else:
collector = solver.FirstSolutionCollector()
# Add decision vars to collector
collector.Add(assignments_flat)
monitor = pywrapcp.SearchMonitor(solver)
monitor.RestartSearch()
# Set time limit if given
if solver_option == 'optimise' and time_limit != None:
print('time_limit', time_limit)
solver_time_limit = solver.TimeLimit(time_limit * 60 * 1000)
# Solve appropriately
if solver_option == 'optimal':
collector.AddObjective(total_cost)
status = solver.Solve(db, [objective, collector, monitor])
elif solver_option == 'optimise' and time_limit != None:
collector.AddObjective(total_cost)
status = solver.Solve(
db, [objective, collector, solver_time_limit, monitor])
else:
status = solver.Solve(db, [collector])
print("Time:", solver.WallTime(), "ms")
print('status', status)
# If solution found, collect all assignments
if status:
solution_by_task = {}
solution_by_worker = {}
for i in range(num_workers):
for j in range(num_tasks):
if collector.Value(0, assignments[i][j]) == 1:
worker_task = assignments_ref[i][j]
# Group solution by worker and task
if worker_task.task.id in solution_by_task:
solution_by_task[worker_task.task.id] = [
*solution_by_task[worker_task.task.id],
worker_task.worker.id
]
else:
solution_by_task[worker_task.task.id] = [
worker_task.worker.id
]
if worker_task.worker.id in solution_by_worker:
solution_by_worker[worker_task.worker.id] = [
*solution_by_worker[worker_task.worker.id],
worker_task.task.id
]
else:
solution_by_worker[worker_task.worker.id] = [
worker_task.task.id
]
if solver_option == 'optimal' or (solver_option == 'optimise'
and time_limit != None):
objective_value = collector.ObjectiveValue(0)
else:
objective_value = get_non_optimised_cost(cost_matrix,
solution_by_task)
return {
"status": status,
"solutionByTask": solution_by_task,
"solutionByWorker": solution_by_worker,
"objectiveValue": objective_value
}
return {
"status": status,
"solutionByTask": None,
"solutionByWorker": None,
"objectiveValue": None
}
""" This script runs a matrix workload that solves a least squares problem using a series of matrix multiplications. """ import subprocess import utils CORES_PER_WORKER = 8 # Figure out the public hostname of the machine. get_hostname_command = "curl -s http://169.254.169.254/latest/meta-data/public-hostname" MASTER_HOSTNAME = subprocess.check_output(get_hostname_command, shell=True) print "Running job with master", MASTER_HOSTNAME workers = utils.get_workers() total_cores = len(workers) * CORES_PER_WORKER # Compute the parameters for the experiment. # Increasing the number of rows increases the CPU time. total_rows_values = [1024 * 1024, 2 * 1024 * 1024] # This is basically the number of tasks; increase this for more tasks. num_row_blocks_values = [total_cores, total_cores * 2] # Reducing this will reduce computation by reduction^2 cols_per_block = 4096 # This is the number of times the shuffle stage will happen. # To use less memory, reduce this number, and instead increase # num_repeats (there's one RDD stored in memory for each shuffle # block). num_col_blocks = 5
This script runs jobs that process the same amount of data, but use different
numbers of tasks to do so.
Each job reads data from in-memory, sorts the data (saving intermediate shuffle
data in-memory) and stores the out in-memory.
"""
import os
import subprocess
import time
import utils
MEGABYTES_PER_GIGABYTE = 1024
slaves = utils.get_workers()
print "Running experiment assuming slaves %s" % slaves
num_machines = len(slaves)
values_per_key = 8
num_shuffles = 5
base_num_tasks = num_machines * 8
num_tasks_multipliers = [8, 4]
target_total_data_gb = num_machines * 0.5
for num_tasks_multiplier in num_tasks_multipliers:
num_tasks = base_num_tasks * num_tasks_multiplier
total_num_items = target_total_data_gb / (4.9 + values_per_key * 1.92) * (
64 * 4000000)
