def test_cost_without_checkpoint(self):
sequence = [4, 10]
handler = rqs.LogDataCost(sequence)
cost = rqs.ClusterCosts(1, 0, 0)
self.assertEqual(handler.compute_cost([3], cost), 4)
self.assertEqual(handler.compute_cost([7], cost), 14)
cost = rqs.ClusterCosts(1, 1, 0)
self.assertEqual(handler.compute_cost([3], cost), 7)
self.assertEqual(handler.compute_cost([7], cost), 25)
cost = rqs.ClusterCosts(1, 1, 1)
self.assertEqual(handler.compute_cost([3], cost), 8)
self.assertEqual(handler.compute_cost([7], cost), 27)
def test_cost_with_checkpoint(self):
sequence = [(4, 1), (6, 0)]
handler = rqs.LogDataCost(sequence)
cost = rqs.ClusterCosts(1, 0, 0)
self.assertEqual(handler.compute_cost([3], cost), 4)
self.assertEqual(handler.compute_cost([7], cost), 10)
cost = rqs.ClusterCosts(1, 1, 0)
self.assertEqual(handler.compute_cost([3], cost), 7)
self.assertEqual(handler.compute_cost([7], cost), 17)
cost = rqs.ClusterCosts(1, 1, 1)
self.assertEqual(handler.compute_cost([3], cost), 8)
self.assertEqual(handler.compute_cost([7], cost), 19)
def test_sequence_cost(self):
wl = rqs.ResourceEstimator([5] * 101)
sequence = wl.compute_request_sequence()
cost = wl.compute_sequence_cost(sequence, [1, 2, 3])
self.assertEqual(cost[0], 7)
cost = rqs.ClusterCosts(0, 1, 0)
sequence = wl.compute_request_sequence(cluster_cost=cost)
cost = wl.compute_sequence_cost(sequence, [1, 2, 3], cluster_cost=cost)
self.assertEqual(cost[0], 2)
def test_system_models(self):
# test the Cloud model (alpha 1 beta 0 gamma 0)
history = np.loadtxt("examples/logs/truncnorm.in", delimiter=' ')
wl = rqs.ResourceEstimator(history)
sequence = wl.compute_request_sequence(cluster_cost=rqs.ClusterCosts(
reservation_cost=1, utilization_cost=0, deploy_cost=0))
self.assertTrue(abs(sequence[0][0] - 10.8) < 0.1)
params = rqs.ResourceParameters()
params.interpolation_model = [
rqs.DistInterpolation(list_of_distr=[norm], discretization=100)
]
wl = rqs.ResourceEstimator(history, params=params)
sequence = wl.compute_request_sequence(cluster_cost=rqs.ClusterCosts(
reservation_cost=1, utilization_cost=0, deploy_cost=0))
self.assertTrue(abs(sequence[0][0] - 10.8) < 0.1)
history = np.loadtxt("examples/logs/CT_eye_segmentation.log",
delimiter=' ')
wl = rqs.ResourceEstimator(history)
sequence = wl.compute_request_sequence(cluster_cost=rqs.ClusterCosts(
reservation_cost=1, utilization_cost=0, deploy_cost=0))
self.assertTrue(abs(sequence[0][0] / 3600 - 22.4) < 0.1)
def default_sequence(training, history, interpolation, discretization,
stype):
cl = rqs.ClusterCosts(1, 1, 0)
training = np.append(training, max(history))
params = rqs.ResourceParameters()
params.interpolation_model = interpolation,
params.resource_discretization = discretization,
params.CR_strategy = stype
wl = rqs.ResourceEstimator(training, params=params)
sequence = wl.compute_request_sequence(cluster_cost=cl)
cost = wl.compute_sequence_cost(sequence, history, cluster_cost=cl)
return cost
if __name__ == '__main__':
if len(sys.argv) < 3:
print("Usage: %s log_file training_size" % (sys.argv[0]))
exit()
file_name = sys.argv[1]
train_size = int(sys.argv[2])
data = np.loadtxt(file_name, delimiter=' ')
assert (len(data) > train_size), "Training size exceeds the total dataset"
training = list(data[:train_size]) + [max(data)]
# compute the requests based on the entire data
wl = rqs.ResourceEstimator(data)
cl = rqs.ClusterCosts(1, 0, 0)
sequence = wl.compute_request_sequence(cluster_cost=cl)
# cost value will be used as reference as optimal
cost_opt = wl.compute_sequence_cost(sequence, data, cluster_cost=cl)
print("Request sequence based on the entire dataset: %s Cost %.2f" %
(sequence, cost_opt))
# compute the requests based on the training data
wl = rqs.ResourceEstimator(training)
sequence = wl.compute_request_sequence(cluster_cost=cl)
cost = wl.compute_sequence_cost(sequence, data, cluster_cost=cl)
print("Request sequence based on training: %s\n"\
"Sequence cost: %.2f (within %.2f%% of optimal)" % (
sequence, cost, (cost-cost_opt)*100/cost_opt))
import sys
sys.path.append("..")
import iSBatch as rqs
import numpy as np
if __name__ == '__main__':
if len(sys.argv) < 3:
print("Usage: %s walltime_log_file memory_file" % (sys.argv[0]))
exit()
file_name = sys.argv[1]
history = np.loadtxt(file_name, delimiter=' ')
memory_footprint = np.loadtxt(sys.argv[2], delimiter=',')
# set the cluster cost model
params = rqs.ResourceParameters()
params.CR_strategy = rqs.CRStrategy.AlwaysCheckpoint
check_model = rqs.StaticCheckpointMemoryModel(
checkpoint_cost=np.max(memory_footprint),
restart_cost=np.max(memory_footprint))
cl_cost = rqs.ClusterCosts(checkpoint_memory_model=check_model)
wl = rqs.ResourceEstimator(history, params=params)
sequence = wl.compute_request_sequence(cluster_cost=cl_cost)
print("Request sequence (static checkpoint): %s" % (sequence))
check_model = rqs.DynamicCheckpointMemoryModel(memory_footprint)
cl_cost = rqs.ClusterCosts(checkpoint_memory_model=check_model)
sequence = wl.compute_request_sequence(cluster_cost=cl_cost)
print("Request sequence (dynamic checkpoint): %s" % (sequence))