Commit d5b5d8c1 authored by Valentin Reis's avatar Valentin Reis

[fix] The epsilon-greedy bandit strategy works.

parent 8c544aef
......@@ -3,6 +3,7 @@ from __future__ import print_function
import logging
import itertools
import numpy
import math
logger = logging.getLogger('nrm')
......@@ -57,22 +58,23 @@ class DiscretizedPowerActuator(object):
"""Actuator in charge of power control via discretization."""
def __init__(self, sm, lowerboundwatts, n):
def __init__(self, sm, lowerboundwatts, k):
self.sensor_manager = sm
self.lowerboundwatts = lowerboundwatts # the minimal cpu wattage
self.n = n # the number of arms
self.k = k # the number of arms
def available_actions(self):
actions = []
pl = self.sensor_manager.get_powerlimits()"BanditPowerActuator: power limits %r", pl)
maxW = int(pl[k]['maxW'])
maxW = int(pl[[k for k,i in pl.items()][0]]['maxW'])
if maxW > self.lowerboundwatts:
logger.error( "BanditPowerActuator: The provided power lowerbound is"\
"lower than the available maximum CPU wattage.")
arms = [self.lowerboundwatts + (float(a)*rangeW/float(n)) for a in range(1,n+1)]
arms = [self.lowerboundwatts + (float(a)*rangeW/float(self.k)) for a in range(1,self.k+1)]"BanditPowerActuator: discretized power limits: %r:", arms)
actions = [Action(target,a,target-a) for a in arms]
actions = [Action([k for k,i in pl.items()][0],int(math.floor(a)),0) for a in arms]
def execute(self, action):
......@@ -84,22 +86,21 @@ class DiscretizedPowerActuator(object):
class BasicPowerLoss(object):
def __init__(self, a, b, power_min, power_max, progress_min, progress_max):
assert(a < b)
def __init__(self, a, b, power_min=100000000, power_max=0, progress_min=1000000000, progress_max=0):
self.a = a
self.b = b
self.power_min = 100000000
self.power_max = 0
self.progress_min = 1000000000
self.progress_max = 0
def perf(self,progress,power):
if power>power_max: power_max = power
if power<power_min: power_min = power
if progress>progress_max: progress_max = progress
if progress<progress_min: progress_min = progress
return((self.a*(power-power_min)/(power_max-power_min)) +
self.power_min = power_min
self.power_max = power_max
self.progress_min = progress_min
self.progress_max = progress_max
def loss(self,progress,power):
if power>self.power_max: self.power_max = power
if power<self.power_min: self.power_min = power
if progress>self.progress_max: self.progress_max = progress
if progress<self.progress_min: self.progress_min = progress
return((self.a*(power-self.power_min)/(max(0.001, self.power_max-self.power_min))) +
class EpsGreedyBandit(object):
"""Epsilon greedy bandit. Actions in O,..,k-1."""
......@@ -113,21 +114,25 @@ class EpsGreedyBandit(object):
def next(self, loss):
assert(loss >= 0)
if self.a:
#"NEXT0!! Arriving with self.n :%s" %str(self.n))
#"NEXT0!! Arriving with self.a :%s" %str(self.a))
if self.a!=None:
self.n=self.n+1"Bandit: the total plays are:%s" %str(self.plays))"Bandit: the estimated losses are:%s" %str([l/p for l,p in zip(self.losses,self.plays)]))
if self.n <= self.k:
self.a = self.n-1
if numpy.random.binomial(1,self.epsilon) == 1:
self.a=numpy.argmin([l/float(n) for l,n in zip(self.losses,self.plays)])
class BanditController(object):
......@@ -135,24 +140,26 @@ class BanditController(object):
def __init__(self, actuators, initialization_rounds=20, exploration=0.2, enforce=None):
self.actuators = actuators
self.initialization_rounds = 20
self.actions = itertools.product(*[act.available_actions() for a in actuators])
self.initialization_rounds = initialization_rounds
self.actions = [a for a in itertools.product(*[act.available_actions() for act in actuators])]
self.loss = BasicPowerLoss(1,-1)
self.bandit = EpsGreedyBandit(exploration,len(self.actions))
if enforce:
if enforce!=None:
def planify(self, target, machineinfo, applications):
"""Plan the next action for the control loop."""
total_progress = sum([a.progress for a in applications.values()])
total_power = float(machineinfo['energy']['power']['total'])
total_power = float(machineinfo['energy']['power']['p0'])"Controller:MACHINEINFO %s." %(str(machineinfo)))"Controller: Reading progress %s and power %s." %(total_progress,total_power))
loss = self.loss(progress=total_progress,power=total_power)
loss = self.loss.loss(progress=total_progress,power=total_power)"Controller: Computing loss %s." %loss)
if self.enforce:
if self.enforce!=None:"Controller: enforced action.")
if self.n>self.initialization_rounds:
......@@ -160,9 +167,9 @@ class BanditController(object):
else:"Controller: estimating max power/max progress ranges.")
a=self.n % k
a=self.n % len(self.actions)
action = self.actions[a]"Controller: playing arm id %a (powercap '%r')." %(a,action.command))"Controller: playing arm id %s (powercap '%s')." %(str(a),str([act.command for act in list(action)])))
def execute(self, actions, actuators):
......@@ -170,7 +177,7 @@ class BanditController(object):
for action, actuator in zip(actions,actuators):
def update(self, action, actuator):
def update(self, actions, actuators):
"""Update tracking across the board to reflect the last action."""
for action, actuator in zip(actions,actuators):
......@@ -127,11 +127,9 @@ class Daemon(object):"sending sensor message: %r", msg)
def do_control(self):
plan = self.controller.planify(, self.machine_info, self.application_manager.applications)
actions, actuators = plan
if action:
self.controller.execute(actions, actuators)
self.controller.update(actions, actuators)
actions, actuators = self.controller.planify(, self.machine_info, self.application_manager.applications)
self.controller.execute(actions, actuators)
self.controller.update(actions, actuators)
def do_signal(self, signum, frame):
if signum == signal.SIGINT:
......@@ -222,7 +220,7 @@ class Daemon(object):
self.application_manager = ApplicationManager()
self.sensor_manager = SensorManager()
# aa = ApplicationActuator(self.application_manager, self.downstream_pub)
pa = DiscretizedPowerActuator(self.sensor_manager,100,4)
pa = DiscretizedPowerActuator(self.sensor_manager,lowerboundwatts=100,k=4)
self.controller = BanditController([pa])
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment