class LRU(page_replacement_algorithm):
def __init__(self, N):
self.T = []
self.N = N
self.disk = CacheLinkedList(N)
self.unique = {}
self.unique_cnt = 0
self.pollution_dat_x = []
self.pollution_dat_y = []
self.time = 0
def get_N(self) :
return self.N
def getWeights(self):
# return np.array([self. X, self.Y1, self.Y2,self.pollution_dat_x,self.pollution_dat_y ]).T
return np.array([self.pollution_dat_x,self.pollution_dat_y ]).T
def getStats(self):
d={}
d['pollution'] = np.array([self.pollution_dat_x, self.pollution_dat_y ]).T
return d
def request(self,page) :
self.time = self.time + 1
page_fault = False
if self.disk.inDisk(page) :
self.disk.moveBack(page)
else :
if self.disk.size() == self.N :
## Remove LRU page
lru = self.disk.getFront()
self.disk.delete(lru)
# Add page to the MRU position
self.disk.add(page)
page_fault = True
if page_fault :
self.unique_cnt += 1
self.unique[page] = self.unique_cnt
if self.time % self.N == 0:
pollution = 0
for pg in self.disk:
if self.unique_cnt - self.unique[pg] >= 2*self.N:
pollution += 1
self.pollution_dat_x.append(self.time)
self.pollution_dat_y.append(100*pollution / self.N)
return page_fault
def get_data(self):
return [self.disk.get_data()]
def get_list_labels(self) :
return ['L']
def __init__(self, N, visualization=True):
self.N = N
self.Cache = ClockDT2(N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
## Config variables
self.learningRate = 0.45
self.error_discount_rate = (0.005)**(1.0 / N)
##
self.policy = 0
self.evictionTime = {}
self.policyUsed = {}
self.weightsUsed = {}
## Accounting variables
self.time = 0
self.W = np.array([.5, .5], dtype=np.float32)
self.Visualization = visualization
self.X = np.array([], dtype=np.int32)
self.Y1 = np.array([])
self.Y2 = np.array([])
###
self.q = Queue.Queue()
self.sum = 0
self.NewPages = []
self.TR = {}
def __init__(self, N, visualization=True):
self.N = N
self.H = N
self.CacheRecency = CacheLinkedList(N)
self.freq = {}
self.PQ = []
self.Hist = CacheLinkedList(self.H)
## Config variables
self.error_discount_rate = (0.005)**(1.0 / N)
self.learning_rate = 0.5
##
self.evictionTime = {}
self.policyUsed = {}
## Accounting variables
self.time = 0
self.W = np.array([.5, .5], dtype=np.float32)
self.Visualization = visualization
self.X = []
self.Y1 = []
self.Y2 = []
self.unique = {}
self.unique_cnt = 0
self.pollution_dat_x = []
self.pollution_dat_y = []
def __init__(self, N):
self.T = []
self.N = N
self.disk = CacheLinkedList(N)
self.unique = {}
self.unique_cnt = 0
self.pollution_dat_x = []
self.pollution_dat_y = []
self.time = 0
def __init__(self, param):
assert 'cache_size' in param
self.N = param['cache_size']
self.T = []
self.disk = CacheLinkedList(self.N)
self.unique = {}
self.unique_cnt = 0
self.pollution_dat_x = []
self.pollution_dat_y = []
self.time = 0
def __init__(self, N, visualization=True):
self.N = N
self.CacheRecency = CacheLinkedList(N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
self.Hist3 = CacheLinkedList(N)
## Config variables
self.error_discount_rate = (0.005)**(1.0 / N)
self.learning_rate = 0.5
self.counter = 0
##
self.evictionTime = {}
## Accounting variables
self.time = 0
self.W = np.array([.33, .33, .33], dtype=np.float32)
self.Visualization = visualization
self.X = []
self.Y1 = []
self.Y2 = []
self.Y3 = []
self.gamma = 0.05 # uniform distribution mixture parameter
self.q_used = {}
self.unique = {}
self.unique_cnt = 0
self.pollution_dat_x = []
self.pollution_dat_y = []
def __init__(self, param):
assert 'cache_size' in param
assert 'history_size_multiple' in param
self.N = int(param['cache_size'])
self.H = int(self.N * int(param['history_size_multiple']) / 2)
self.learning_rate = float(
param['learning_rate']) if 'learning_rate' in param else 0
self.Visualization = 'visualization' in param and bool(
param['visualization'])
self.CacheRecency = CacheLinkedList(self.N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(self.H)
self.Hist2 = CacheLinkedList(self.H)
## Accounting variables
self.time = 0
self.W = np.array([.5, .5], dtype=np.float32)
self.X = []
self.Y1 = []
self.Y2 = []
self.unique = {}
self.unique_cnt = 0
self.pollution_dat_x = []
self.pollution_dat_y = []
def __init__(self, N):
self.N = N
self.CacheRecency = CacheLinkedList(N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
## Config variables
self.epsilon = 0.05
self.error_discount_rate = (0.005)**(1.0 / N)
self.policy_space_size = 4
self.Gamma = 0.5
self.minWeight = 0.01
##
self.evictionTime = {}
self.policyUsed = {}
self.weightsUsed = {}
self.qUsed = {}
self.freq = {}
## Accounting variables
self.time = 0
self.unif = self.Gamma * np.ones(
self.policy_space_size, dtype=np.float64) / self.policy_space_size
self.W = np.ones(self.policy_space_size,
dtype=np.float64) / self.policy_space_size
self.X = np.array([], dtype=np.int32)
self.Y = np.array([])
###
self.q = Queue.Queue()
self.sum = 0
self.NewPages = []
def __init__(self, N, visualization=True):
self.N = N
self.freq = {}
self.PQ = []
self.Cache = CacheLinkedList(N)
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
def __init__(self, param):
assert 'cache_size' in param
self.N = int(param['cache_size'])
self.H = int(self.N * int(param['history_size_multiple'])
) if 'history_size_multiple' in param else self.N
self.learning_rate = float(
param['learning_rate']) if 'learning_rate' in param else 0
self.initial_weight = float(
param['initial_weight']) if 'initial_weight' in param else 0.5
self.discount_rate = float(
param['discount_rate']) if 'discount_rate' in param else 1
self.Visualization = 'visualize' in param and bool(param['visualize'])
#self.discount_rate = 0.005**(1/self.N)
np.random.seed(123)
self.CacheRecency = CacheLinkedList(self.N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(self.H)
self.Hist2 = CacheLinkedList(self.H)
## Accounting variables
self.time = 0
self.W = np.array([self.initial_weight, 1 - self.initial_weight],
dtype=np.float32)
self.X = []
self.Y1 = []
self.Y2 = []
self.eTime = {}
self.unique = {}
self.unique_cnt = 0
self.pollution_dat_x = []
self.pollution_dat_y = []
self.pollution_dat_y_val = 0
self.pollution_dat_y_sum = []
self.pollution = 0
self.learning_rates = []
self.info = {
'lru_misses': 0,
'lfu_misses': 0,
'lru_count': 0,
'lfu_count': 0,
}
def __init__(self, N):
self.N = N
self.CacheRecency = CacheLinkedList(N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
## Config variables
self.epsilon = 0.90
self.error_discount_rate = (0.005)**(1.0 / N)
self.Gamma = 0.05
##
self.policy = 0
self.evictionTime = {}
self.policyUsed = {}
self.weightsUsed = {}
self.qUsed = {}
## Accounting variables
self.time = 0
self.W = np.array([.5, .5], dtype=np.float32)
self.X = np.array([], dtype=np.int32)
self.Y1 = np.array([])
self.Y2 = np.array([])
###
self.q = Queue.Queue()
self.sum = 0
self.NewPages = []
self.TR = {}
class LeCaR10(page_replacement_algorithm):
# def __init__(self, N, visualization = True):
def __init__(self, param):
assert 'cache_size' in param
# assert 'history_size_multiple' in param
self.N = int(param['cache_size'])
self.H = int(self.N * int(param['history_size_multiple'])) if 'history_size_multiple' in param else self.N
self.learning_rate = float(param['learning_rate']) if 'learning_rate' in param else 0
self.discount_rate = float(param['discount_rate']) if 'discount_rate' in param else 1
# self.discount_rate = (float(param['discount_rate']) if 'discount_rate' in param else 0) ** (1/self.N)
# self.discount_rate = 0.05**(1/self.N)
self.Visualization = 'visualize' in param and bool(param['visualize'])
self.lamb = 0.05
self.learning_rate = 0.1
self.learning_rate_lfu= 0.1
self.learning_rate_lru= 0.1
self.CacheRecency = CacheLinkedList(self.N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(self.H)
self.Hist2 = CacheLinkedList(self.H)
np.random.seed(123)
self.PageCount = 0
self.CacheHit = 0
self.PreviousHR = 0.0
self.NewHR = 0.0
self.PreviousChangeInHR = 0.0
self.NewChangeInHR =0.0
self.PreviousLR= 0.45
self.NewLR =0.45
self.CacheHitList = []
self.counter = 0
self.learning_rates = []
self.SampleHR =[]
self.SAMPLE_SIZE = 20 * self.N
self.SampleHitQ = queue.Queue(maxsize= self.SAMPLE_SIZE)
self.SampleCacheHit = 0
## Accounting variables
self.time = 0
self.W = np.array([.5,.5], dtype=np.float32)
self.PreviousW = np.array([.5,.5], dtype=np.float32)
self.NewW = np.array([.5,.5], dtype=np.float32)
self.qUsed = {}
self.eTime = {}
self.X = []
self.Y1 = []
self.Y2 = []
self.unique = {}
self.unique_cnt = 0
self.pollution_dat_x = []
self.pollution_dat_y = []
self.pollution_dat_y_val = 0
self.pollution_dat_y_sum = []
self.pollution =0
def __contains__(self, q):
return q in self.CacheRecency
def get_N(self) :
return self.N
def visualize(self, ax_w, ax_h, averaging_window_size):
lbl = []
if self.Visualization:
X = np.array(self.X)
Y1 = np.array(self.Y1)
Y2 = np.array(self.Y2)
ax_w.set_xlim(np.min(X), np.max(X))
ax_h.set_xlim(np.min(X), np.max(X))
ax_w.plot(X,Y1, 'y-', label='W_lru', linewidth=2)
ax_w.plot(X,Y2, 'b-', label='W_lfu', linewidth=1)
#ax_h.plot(self.pollution_dat_x,self.pollution_dat_y, 'g-', label='hoarding',linewidth=3)
#ax_h.plot(self.pollution_dat_x,self.pollution_dat_y, 'k-', linewidth=3)
ax_h.set_ylabel('Hoarding')
ax_w.legend(loc=" upper right")
ax_w.set_title('LeCaR - Adaptive LR')
pollution_sums = self.getPollutions()
temp = np.append(np.zeros(averaging_window_size), pollution_sums[:-averaging_window_size])
pollutionrate = (pollution_sums-temp) / averaging_window_size
ax_h.set_xlim(0, len(pollutionrate))
ax_h.plot(range(len(pollutionrate)), pollutionrate, 'k-', linewidth=3)
# lbl.append(l1)
# lbl.append(l2)
# lbl.append(l3)
return lbl
def getWeights(self):
return np.array([self. X, self.Y1, self.Y2,self.pollution_dat_x,self.pollution_dat_y_sum ]).T
# return np.array([self.pollution_dat_x,self.pollution_dat_y ]).T
def getPollutions(self):
return self.pollution_dat_y_sum
def getLearningRates(self):
return self.learning_rates
def getStats(self):
d={}
d['weights'] = np.array([self. X, self.Y1, self.Y2]).T
d['pollution'] = np.array([self.pollution_dat_x, self.pollution_dat_y ]).T
return d
##############################################################
## There was a page hit to 'page'. Update the data structures
##############################################################
def pageHitUpdate(self, page):
assert page in self.CacheRecency and page in self.freq
self.CacheRecency.moveBack(page)
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page],page))
##########################################
## Add a page to cache using policy 'poly'
##########################################
def addToCache(self, page):
self.CacheRecency.add(page)
if page not in self.freq :
self.freq[page] = 0
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page],page))
def getHeapMin(self):
while self.PQ[0][1] not in self.CacheRecency or self.freq[self.PQ[0][1]] != self.PQ[0][0] :
heapq.heappop(self.PQ)
return self.PQ[0][1]
######################
## Get LFU or LFU page
######################
def selectEvictPage(self, policy):
r = self.CacheRecency.getFront()
f = self.getHeapMin()
pageToEvit,policyUsed = None, None
if r == f :
pageToEvit,policyUsed = r,-1
elif policy == 0:
pageToEvit,policyUsed = r,0
elif policy == 1:
pageToEvit,policyUsed = f,1
return pageToEvit, policyUsed
def evictPage(self, pg):
assert pg in self.CacheRecency
self.CacheRecency.delete(pg)
def getQ(self):
lamb = 0.05
return (1-lamb)*self.W + lamb/2
############################################
## Choose a page based on the q distribution
############################################
def chooseRandom(self):
# np.random.seed(10)
r = np.random.rand()
# q = self.getQ()
if r < self.W[0] :
return 0
return 1
def addToHistory(self, poly, cacheevict):
histevict = None
if (poly == 0) or (poly==-1 and np.random.rand() <0.5):
if self.Hist1.size() == self.H :
histevict = self.Hist1.getFront()
assert histevict in self.Hist1
self.Hist1.delete(histevict)
self.Hist1.add(cacheevict)
else:
if self.Hist2.size() == self.H :
histevict = self.Hist2.getFront()
assert histevict in self.Hist2
self.Hist2.delete(histevict)
self.Hist2.add(cacheevict)
if histevict is not None :
del self.freq[histevict]
del self.qUsed[histevict]
del self.eTime[histevict]
def updateLearningRates(self,seq_len):
if self.time % (seq_len) == 0:
# print("Inside",self.N/2)
# self.NewHR = np.mean(self.CacheHitList)
self.NewHR = self.CacheHit/ float(seq_len)
self.NewChangeInHR= (self.NewHR -self.PreviousHR)
self.NewCahngeInW = self.W - self.PreviousW
delta_1 = self.NewChangeInHR
delta_2 = self.PreviousChangeInHR
delta = delta_1 * delta_2
delta_lfu = 0
delta_lru = 0
# print( "ChangeinW", abs(self.NewCahngeInW[0]), abs(self.NewCahngeInW[1]))
if self.NewCahngeInW[0] != 0:
delta_lfu = self.NewChangeInHR / (self.NewCahngeInW[0])
if self.NewCahngeInW[1] != 0:
delta_lru = self.NewChangeInHR / (self.NewCahngeInW[1])
# print("Change in HR",self.NewChangeInHR )
# print("LFU Previous and New Weights",self.PreviousW[0], self.W[0])
# print("LRU Previous and New Weights",self.PreviousW[1], self.W[1])
# print("delta", delta_lfu, delta_lru)
# if self.learning_rate * (self.SampleCacheHit/ float(self.SAMPLE_SIZE)) <10**(-5):
# # if self.learning_rate==0 :
# self.learning_rate = 0.1
# else: self.learning_rate = self.learning_rate*2
# elif self.learning_rate * (self.SampleCacheHit/ float(self.SAMPLE_SIZE)) <10**(-6):
# # if self.learning_rate * delta * delta_1 <10**(-5):
# self.learning_rate = 0.45
if delta_lfu< 0 and delta_lfu!=0:
self.learning_rate_lfu = max(self.learning_rate_lfu /2, 0 )
elif delta_lfu >0 and delta_lfu!=0:
self.learning_rate_lfu = min(self.learning_rate_lfu + (0.1* self.learning_rate_lfu), 1 )
if delta_lru< 0 and delta_lru!=0:
self.learning_rate_lru = max(self.learning_rate_lru/2 , 0 )
elif delta_lru >0 and delta_lru!=0:
self.learning_rate_lru = min(self.learning_rate_lru + (0.1* self.learning_rate_lru), 1 )
self.PreviousLR = self.NewLR
self.NewLR = self.learning_rate
self.PreviousHR = self.NewHR
self.PreviousChangeInHR = self.NewChangeInHR
del self.CacheHitList[:]
self.CacheHit = 0
self.PreviousW = self.W
########################################################################################################################################
####REQUEST#############################################################################################################################
########################################################################################################################################
def request(self,page) :
page_fault = False
self.time = self.time + 1
# print(self.PageCount)
###########################
## Clean up
## In case PQ get too large
##########################
if len(self.PQ) > 2*self.N:
newpq = []
for pg in self.CacheRecency:
newpq.append((self.freq[pg],pg))
heapq.heapify(newpq)
self.PQ = newpq
del newpq
#####################
## Visualization data
#####################
if self.Visualization:
self.X.append(self.time)
self.Y1.append(self.W[0])
self.Y2.append(self.W[1])
#####################################################
## Adapt learning rate Here
######################################################
seq_len = self.N
self.updateLearningRates(seq_len)
# if len(self.SampleHR)== 20* self.N:
# del self.SampleHR[0]
if self.SampleHitQ.full():
self.SampleCacheHit -= self.SampleHitQ.get()
##########################
## Process page request
##########################
if page in self.CacheRecency:
page_fault = False
self.CacheHit +=1
self.CacheHitList.append(1)
self.SampleCacheHit += 1
self.SampleHitQ.put(1)
# self.SampleHR.append(1)
self.pageHitUpdate(page)
else :
# updateLearningRates(self,seq_len)
#####################################################
## Learning step: If there is a page fault in history
#####################################################
pageevict = None
self.CacheHitList.append(0)
self.SampleHitQ.put(0)
# self.SampleHR.append(0)
reward = np.array([0,0], dtype=np.float32)
if page in self.Hist1:
pageevict = page
self.Hist1.delete(page)
# reward[0] = -1 / self.qUsed[page] ## punish
reward[0] = -self.discount_rate **( (self.time-self.eTime[page]) ) ## punish
# reward[0] = -1 ## punish
elif page in self.Hist2:
pageevict = page
self.Hist2.delete(page)
# reward[1] = -1 / self.qUsed[page]
reward[1] = -self.discount_rate ** ( (self.time-self.eTime[page]) )
# reward[1] = -1 ## punish
#################
## Update Weights
#################
if pageevict is not None :
# self.W = self.W * np.exp(self.learning_rate * reward )
self.W[0] = self.W[0] * np.exp(self.learning_rate_lfu * reward[0] )
self.W[1] = self.W[1] * np.exp(self.learning_rate_lru * reward[1] )
self.W = self.W / np.sum(self.W)
# self.W[0] = min(1-self.lamb, self.W[0])
# self.W[0] = max(self.lamb, self.W[0])
# self.W[1] = 1 - self.W[0]
# self.PreviousW = self.W
####################
## Remove from Cache
####################
if self.CacheRecency.size() == self.N:
################
## Choose Policy
################
act = self.chooseRandom()
cacheevict,poly = self.selectEvictPage(act)
# self.qUsed[cacheevict] = self.getQ()[poly]
self.qUsed[cacheevict] = self.W[poly]
self.eTime[cacheevict] = self.time
###################
## Remove from Cache and Add to history
###################
self.evictPage(cacheevict)
self.addToHistory(poly, cacheevict)
self.addToCache(page)
page_fault = True
## Count pollution
if page_fault:
self.unique_cnt += 1
self.unique[page] = self.unique_cnt
if self.time % self.N == 0:
self.pollution = 0
for pg in self.CacheRecency:
if self.unique_cnt - self.unique[pg] >= 2*self.N:
self.pollution += 1
self.pollution_dat_x.append(self.time)
self.pollution_dat_y.append(100* self.pollution / self.N)
self.pollution_dat_y_val += 100* self.pollution / self.N
self.pollution_dat_y_sum.append(self.pollution_dat_y_val)
self.learning_rates.append(self.learning_rate)
return page_fault
def get_list_labels(self) :
return ['L']
class LeCaR(page_replacement_algorithm):
# def __init__(self, N, visualization = True):
def __init__(self, param):
assert 'cache_size' in param
self.N = int(param['cache_size'])
self.H = int(self.N * int(param['history_size_multiple'])
) if 'history_size_multiple' in param else self.N
self.discount_rate = float(
param['discount_rate']) if 'discount_rate' in param else 1
self.learning_rate = float(
param['learning_rate']) if 'learning_rate' in param else 0
self.initial_weight = float(
param['initial_weight']) if 'initial_weight' in param else 0.5
self.Visualization = 'visualize' in param and bool(param['visualize'])
self.discount_rate = 0.005**(1 / self.N)
self.CacheRecency = CacheLinkedList(self.N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(self.H)
self.Hist2 = CacheLinkedList(self.H)
np.random.seed(123)
## Accounting variables
self.time = 0
self.eTime = {}
self.W = np.array([self.initial_weight, 1 - self.initial_weight],
dtype=np.float32)
self.X = []
self.Y1 = []
self.Y2 = []
self.unique = {}
self.unique_cnt = 0
self.pollution_dat_x = []
self.pollution_dat_y = []
self.info = {
'lru_misses': 0,
'lfu_misses': 0,
'lru_count': 0,
'lfu_count': 0,
}
def get_N(self):
return self.N
def __contains__(self, q):
return q in self.CacheRecency
def visualize(self, ax):
lbl = []
if self.Visualization:
X = np.array(self.X)
Y1 = np.array(self.Y1)
Y2 = np.array(self.Y2)
# ax = plt.subplot(2,1,1)
ax.set_xlim(np.min(X), np.max(X))
# l3, = plt.plot(self.pollution_dat_x,self.pollution_dat_y, 'g-', label='hoarding',linewidth=3)
# l1, = plt.plot(X,Y1, 'y-', label='W_lru',linewidth=2)
# l2, = plt.plot(X,Y2, 'b-', label='W_lfu',linewidth=1)
ax.plot(X, Y1, 'y-', label='W_lru', linewidth=2)
ax.plot(X, Y2, 'b-', label='W_lfu', linewidth=1)
#print "lru_misses = ", self.info['lru_misses']
#print "lru_count = ", self.info['lru_count']
#print "lfu_misses = ", self.info['lfu_misses']
#print "lfu_count = ", self.info['lfu_count']
# lbl.append(l1)
# lbl.append(l2)
# lbl.append(l3)
return lbl
def getWeights(self):
return np.array([
self.X, self.Y1, self.Y2, self.pollution_dat_x,
self.pollution_dat_y
]).T
# return np.array([self.pollution_dat_x,self.pollution_dat_y ]).T
def getStats(self):
d = {}
d['weights'] = np.array([self.X, self.Y1, self.Y2]).T
d['pollution'] = np.array([self.pollution_dat_x,
self.pollution_dat_y]).T
return d
##############################################################
## There was a page hit to 'page'. Update the data structures
##############################################################
def pageHitUpdate(self, page):
assert page in self.CacheRecency and page in self.freq
self.CacheRecency.moveBack(page)
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
##########################################
## Add a page to cache using policy 'poly'
##########################################
def addToCache(self, page):
self.CacheRecency.add(page)
if page not in self.freq:
self.freq[page] = 0
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
def getHeapMin(self):
while self.PQ[0][1] not in self.CacheRecency or self.freq[
self.PQ[0][1]] != self.PQ[0][0]:
heapq.heappop(self.PQ)
return self.PQ[0][1]
######################
## Get LFU or LFU page
######################
def selectEvictPage(self, policy):
r = self.CacheRecency.getFront()
f = self.getHeapMin()
pageToEvit, policyUsed = None, None
#if r == f :
#pageToEvit,policyUsed = r,-1
if policy == 0:
pageToEvit, policyUsed = r, 0
elif policy == 1:
pageToEvit, policyUsed = f, 1
return pageToEvit, policyUsed
def evictPage(self, pg):
assert pg in self.CacheRecency
self.CacheRecency.delete(pg)
def getQ(self):
lamb = 0.05
return (1 - lamb) * self.W + lamb
############################################
## Choose a page based on the q distribution
############################################
def chooseRandom(self):
r = np.random.rand()
if r < self.W[0]:
return 0
return 1
def addToHistory(self, poly, cacheevict):
histevict = None
if (poly == 0) or (poly == -1 and np.random.rand() < 0.5):
if self.Hist1.size() == self.H:
histevict = self.Hist1.getFront()
assert histevict in self.Hist1
self.Hist1.delete(histevict)
self.Hist1.add(cacheevict)
self.info['lru_count'] += 1
else:
if self.Hist2.size() == self.H:
histevict = self.Hist2.getFront()
assert histevict in self.Hist2
self.Hist2.delete(histevict)
self.Hist2.add(cacheevict)
self.info['lfu_count'] += 1
if histevict is not None:
del self.freq[histevict]
del self.eTime[histevict]
########################################################################################################################################
####REQUEST#############################################################################################################################
########################################################################################################################################
def request(self, page):
page_fault = False
self.time = self.time + 1
###########################
## Clean up
## In case PQ get too large
##########################
if len(self.PQ) > 2 * self.N:
newpq = []
for pg in self.CacheRecency:
newpq.append((self.freq[pg], pg))
heapq.heapify(newpq)
self.PQ = newpq
del newpq
#####################
## Visualization data
#####################
if self.Visualization:
self.X.append(self.time)
self.Y1.append(self.W[0])
self.Y2.append(self.W[1])
##########################
## Process page request
##########################
if page in self.CacheRecency:
page_fault = False
self.pageHitUpdate(page)
else:
#####################################################
## Learning step: If there is a page fault in history
#####################################################
pageevict = None
reward = np.array([0, 0], dtype=np.float32)
if page in self.Hist1:
pageevict = page
self.Hist1.delete(page)
#reward[0] = -1
reward[0] = -self.discount_rate**(self.time - self.eTime[page])
self.info['lru_misses'] += 1
elif page in self.Hist2:
pageevict = page
self.Hist2.delete(page)
#reward[1] = -1
reward[1] = -self.discount_rate**(self.time - self.eTime[page])
self.info['lfu_misses'] += 1
#################
## Update Weights
#################
if pageevict is not None:
self.W = self.W * np.exp(self.learning_rate * reward)
self.W = self.W / np.sum(self.W)
####################
## Remove from Cache
####################
if self.CacheRecency.size() == self.N:
################
## Choose Policy
################
act = self.chooseRandom()
cacheevict, poly = self.selectEvictPage(act)
###################
## Remove from Cache and Add to history
###################
self.eTime[cacheevict] = self.time
self.evictPage(cacheevict)
self.addToHistory(poly, cacheevict)
self.addToCache(page)
page_fault = True
## Count pollution
# if page_fault:
# self.unique_cnt += 1
# self.unique[page] = self.unique_cnt
#
# if self.time % self.N == 0:
# pollution = 0
# for pg in self.CacheRecency:
# if self.unique_cnt - self.unique[pg] >= 2*self.N:
# pollution += 1
#
# self.pollution_dat_x.append(self.time)
# self.pollution_dat_y.append(100* pollution / self.N)
return page_fault
def get_list_labels(self):
return ['L']
def __init__(self, N):
self.N = N
self.CacheRecency = CacheLinkedList(N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
## Config variables
self.epsilon = 0.90
self.error_discount_rate = (0.005)**(1.0 / N)
##
self.policy = 0
self.evictionTime = {}
self.policyUsed = {}
self.pUsed = {}
self.param = {}
## Accounting variables
self.time = 0
###
self.q = Queue.Queue()
self.sum = 0
self.NewPages = []
self.c_hits = 0
self.h_miss = 0
self.learning = True
self.X = tf.placeholder(dtype=tf.int32, shape=[None, 2])
self.P = tf.placeholder(dtype=tf.float32, shape=[None, 1])
self.R = tf.placeholder(dtype=tf.float32, shape=[None])
self.F = tf.placeholder(dtype=tf.float32, shape=[None])
self.W = tf.Variable(tf.random_uniform([2 * self.N]))
# self.predict = tf.sigmoid(tf.matmul(self.X, self.W))
# self.predict = tf.sigmoid(tf.slice(self.W, self.X[0,0],[1]) + tf.slice(self.W, self.X[0,1]+self.N,[1]))
idx1 = tf.slice(self.X, [0, 0], [-1, 1])
idx2 = tf.slice(self.X, [0, 1], [-1, 1])
self.w1 = tf.slice(self.W, idx1[0], [1])
self.w2 = tf.slice(self.W, idx2[0], [1])
self.predict = tf.sigmoid(self.w1 + self.w2)
self.cost = tf.reduce_sum(self.R * tf.log(self.predict) +
self.F * tf.log(1 - self.predict))
learning_rate = 0.1
self.optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(self.cost)
##################################
self.X_holder = []
self.P_holder = []
self.R_holder = []
self.F_holder = []
self.train_batch_size = 5 * self.N
init = tf.global_variables_initializer()
self.sess = tf.Session()
self.sess.run(init)
class OLCR_RAND(page_replacement_algorithm):
def __init__(self, N):
self.N = N
self.CacheRecency = CacheLinkedList(N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
## Config variables
self.epsilon = 0.90
self.error_discount_rate = (0.005)**(1.0 / N)
##
self.policy = 0
self.evictionTime = {}
self.policyUsed = {}
self.pUsed = {}
self.param = {}
## Accounting variables
self.time = 0
###
self.q = Queue.Queue()
self.sum = 0
self.NewPages = []
self.c_hits = 0
self.h_miss = 0
self.learning = True
self.X = tf.placeholder(dtype=tf.int32, shape=[None, 2])
self.P = tf.placeholder(dtype=tf.float32, shape=[None, 1])
self.R = tf.placeholder(dtype=tf.float32, shape=[None])
self.F = tf.placeholder(dtype=tf.float32, shape=[None])
self.W = tf.Variable(tf.random_uniform([2 * self.N]))
# self.predict = tf.sigmoid(tf.matmul(self.X, self.W))
# self.predict = tf.sigmoid(tf.slice(self.W, self.X[0,0],[1]) + tf.slice(self.W, self.X[0,1]+self.N,[1]))
idx1 = tf.slice(self.X, [0, 0], [-1, 1])
idx2 = tf.slice(self.X, [0, 1], [-1, 1])
self.w1 = tf.slice(self.W, idx1[0], [1])
self.w2 = tf.slice(self.W, idx2[0], [1])
self.predict = tf.sigmoid(self.w1 + self.w2)
self.cost = tf.reduce_sum(self.R * tf.log(self.predict) +
self.F * tf.log(1 - self.predict))
learning_rate = 0.1
self.optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(self.cost)
##################################
self.X_holder = []
self.P_holder = []
self.R_holder = []
self.F_holder = []
self.train_batch_size = 5 * self.N
init = tf.global_variables_initializer()
self.sess = tf.Session()
self.sess.run(init)
def get_N(self):
return self.N
def visualize(self, plt):
return []
##############################################################
## There was a page hit to 'page'. Update the data structures
##############################################################
def pageHitUpdate(self, page):
assert page in self.CacheRecency and page in self.freq
self.CacheRecency.moveBack(page)
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
##########################################
## Add a page to cache using policy 'poly'
##########################################
def addToCache(self, page):
self.CacheRecency.add(page)
if page not in self.freq:
self.freq[page] = 0
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
def getHeapMin(self):
if len(self.PQ) < self.N:
print self.PQ
assert len(self.PQ) >= self.N, 'PQ should be full %d' % len(self.PQ)
while self.PQ[0][1] not in self.CacheRecency or self.freq[
self.PQ[0][1]] != self.PQ[0][0]:
heapq.heappop(self.PQ)
return self.PQ[0][1]
######################
## Get LFU or LFU page
## return page, poly
######################
def selectEvictPage(self, P):
assert P >= 0 and P <= 1
if np.random.rand() < P:
return self.CacheRecency.getFront(), 0
else:
return self.getHeapMin(), 1
def evictPage(self, pg):
assert pg in self.CacheRecency
self.CacheRecency.delete(pg)
############################################
## Choose a page based on the q distribution
############################################
def chooseRandom(self):
r = np.random.rand()
if r < self.W[0]:
return 0
return 1
def addToHistory(self, poly, cacheevict):
histevict = None
if (poly == 0) or (poly == -1 and np.random.rand() < 0.5):
if self.Hist1.size() == self.N:
histevict = self.Hist1.getFront()
assert histevict in self.Hist1
self.Hist1.delete(histevict)
self.Hist1.add(cacheevict)
else:
if self.Hist2.size() == self.N:
histevict = self.Hist2.getFront()
assert histevict in self.Hist2
self.Hist2.delete(histevict)
self.Hist2.add(cacheevict)
if histevict is not None:
del self.evictionTime[histevict]
del self.policyUsed[histevict]
del self.freq[histevict]
del self.pUsed[histevict]
del self.param[histevict]
########################################################################################################################################
####REQUEST#############################################################################################################################
########################################################################################################################################
def request(self, page):
page_fault = False
self.time = self.time + 1
###########################
## Clean up
## In case PQ get too large
##########################
if len(self.PQ) > 2 * self.N:
newpq = []
for pg in self.CacheRecency:
newpq.append((self.freq[pg], pg))
heapq.heapify(newpq)
self.PQ = newpq
del newpq
page_outcome = -1
##########################
## Process page request
##########################
if page in self.CacheRecency:
page_fault = False
self.pageHitUpdate(page)
page_outcome = 1
else:
#####################################################
## Learning step: If there is a page fault in history
#####################################################
if page in self.Hist1 or page in self.Hist2:
page_outcome = 2
if page in self.Hist1:
self.Hist1.delete(page)
else:
self.Hist2.delete(page)
####################
## Remove from Cache
####################
if self.CacheRecency.size() == self.N:
################
## Choose Policy
################
P = np.random.rand()
cacheevict, poly = self.selectEvictPage(P)
self.policyUsed[cacheevict] = poly
self.evictionTime[cacheevict] = self.time
self.pUsed[cacheevict] = P
self.param[cacheevict] = [self.c_hits, self.h_miss]
###################
## Remove from Cache and Add to history
###################
self.evictPage(cacheevict)
self.addToHistory(poly, cacheevict)
self.addToCache(page)
page_fault = True
self.q.put(page_outcome)
if page_outcome == 1:
self.c_hits += 1
elif page_outcome == 2:
self.h_miss += 1
if self.q.qsize() >= self.N:
temp = self.q.get()
if temp == 1:
self.c_hits -= 1
elif temp == 2:
self.h_miss -= 1
assert self.c_hits >= 0 and self.c_hits < self.N
assert self.h_miss >= 0 and self.h_miss < self.N
return page_fault
def get_list_labels(self):
return ['L']
def __init__(self, N):
self.N = N
self.CacheRecency = CacheLinkedList(N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
## Config variables
self.error_discount_rate = (0.005)**(1.0 / N)
##
self.policy = 0
self.evictionTime = {}
self.policyUsed = {}
self.pUsed = {}
self.param = {}
## Accounting variables
self.time = 0
###
self.q = Queue.Queue()
self.sum = 0
self.NewPages = []
self.c_hits = 0
self.h_miss = 0
self.learning = True
input_units = 2 * self.N
hidden_units = 3
hidden_units2 = 2
output_units = 2
self.X = tf.placeholder(dtype=tf.int32, shape=[None, 2])
self.P = tf.placeholder(dtype=tf.float32, shape=[None, 1])
self.C_r = tf.placeholder(dtype=tf.float32, shape=[None])
self.C_f = tf.placeholder(dtype=tf.float32, shape=[None])
self.W = tf.Variable(tf.random_uniform([input_units, hidden_units]))
self.b = tf.Variable(tf.ones([hidden_units]))
self.W_hidden = tf.Variable(
tf.random_uniform([hidden_units, hidden_units2]))
self.b_hidden = tf.Variable(tf.ones([hidden_units2]))
self.W_hidden2 = tf.Variable(
tf.random_uniform([hidden_units2, output_units]))
self.b_hidden2 = tf.Variable(tf.ones([output_units]))
# self.predict = tf.sigmoid(tf.matmul(self.X, self.W))
# self.predict = tf.sigmoid(tf.slice(self.W, self.X[0,0],[1]) + tf.slice(self.W, self.X[0,1]+self.N,[1]))
idx1 = tf.slice(self.X, [0, 0], [-1, 1])
idx2 = tf.slice(self.X, [0, 1], [-1, 1]) + self.N
## Neuron 1
# self.w1 = tf.slice(self.W, [0, idx1[0]],[1,1])
# self.w2 = tf.slice(self.W, [0, idx2[0]+self.N,[1])
self.w11 = tf.slice(self.W[:, 0], idx1[0], [1])
self.w22 = tf.slice(self.W[:, 1], idx2[0], [1])
self.w13 = tf.slice(self.W[:, 2], idx1[0], [1])
self.w23 = tf.slice(self.W[:, 2], idx2[0], [1])
## Hidden layer
self.neuron1 = tf.sigmoid(self.w11 + self.b[0]) # only x1
self.neuron2 = tf.sigmoid(self.w22 + self.b[1]) # only x2
self.neuron3 = tf.sigmoid(self.w13 + self.w23 + self.b[2]) # x1+x2
temp = tf.concat([[self.neuron1], [self.neuron2], [self.neuron3]], 1)
self.logit1 = tf.contrib.layers.flatten(temp) # 1 x hidden_units
self.logit2 = tf.sigmoid(tf.matmul(self.logit1, self.W_hidden))
self.logit3 = tf.sigmoid(tf.matmul(self.logit2, self.W_hidden2))
# self.predict = tf.nn.softmax(self.logit2)
self.predict = tf.nn.softmax(self.logit3)
# self.cost_r = -tf.reduce_mean(self.C_r * tf.log(1-self.predict[0,0]) + (1-self.C_r) * tf.log(self.predict[0,0]))
# self.cost_f = -tf.reduce_mean(self.C_f * tf.log(1-self.predict[0,1]) + (1-self.C_f) * tf.log(self.predict[0,1]))
# self.cost = self.cost_r + self.cost_f
self.cost = -tf.reduce_mean(self.C_r * tf.log(1 - self.predict[0, 0]) +
self.C_f * tf.log(1 - self.predict[0, 1]))
learning_rate = 0.01
self.optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(self.cost)
##################################
self.X_holder = []
self.P_holder = []
self.R_holder = []
self.F_holder = []
self.train_batch_size = 4 * self.N
init = tf.global_variables_initializer()
self.sess = tf.Session()
self.sess.run(init)
class LaCReME(page_replacement_algorithm):
def __init__(self, N, visualization=True):
self.N = N
self.CacheRecency = CacheLinkedList(N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
## Config variables
self.error_discount_rate = (0.005)**(1.0 / N)
self.learningRate = 0.45
##
self.policy = 0
self.evictionTime = {}
self.policyUsed = {}
self.weightsUsed = {}
## Accounting variables
self.time = 0
self.W = np.array([.5, .5], dtype=np.float32)
self.Visualization = visualization
self.X = []
self.Y1 = []
self.Y2 = []
###
self.q = Queue.Queue()
self.sum = 0
self.NewPages = []
self.TR = {}
def get_N(self):
return self.N
def visualize(self, plt):
lbl = []
if self.Visualization:
X = np.array(self.X)
Y1 = np.array(self.Y1)
Y2 = np.array(self.Y2)
ax = plt.subplot(2, 1, 1)
ax.set_xlim(np.min(X), np.max(X))
l1, = plt.plot(self.X, Y1, 'y-', label='W_lru', linewidth=2)
l2, = plt.plot(self.X, Y2, 'b-', label='W_lfu', linewidth=1)
lbl.append(l1)
lbl.append(l2)
# totaltime = 0
# total2 = 0
# for tc in self.TR:
# if tc is not 'total':
# totaltime += self.TR[tc]
#
# for tc in self.TR:
# if tc is not 'total':
# print '%s = %% %f' % (tc, 100*self.TR[tc] / totaltime)
# total2 += self.TR[tc]
#
# print '%s = %f' % ('total2', total2)
# print '%s = %f' % ('total', self.TR['total'])
#
return lbl
##############################################################
## There was a page hit to 'page'. Update the data structures
##############################################################
def pageHitUpdate(self, page):
assert page in self.CacheRecency and page in self.freq
self.CacheRecency.moveBack(page)
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
##########################################
## Add a page to cache using policy 'poly'
##########################################
def addToCache(self, page):
self.CacheRecency.add(page)
if page not in self.freq:
self.freq[page] = 0
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
def getHeapMin(self):
# if len(self.PQ) < self.N :
# print self.PQ
# assert len(self.PQ) >= self.N, 'PQ should be full %d' % len(self.PQ)
while self.PQ[0][1] not in self.CacheRecency or self.freq[
self.PQ[0][1]] != self.PQ[0][0]:
heapq.heappop(self.PQ)
return self.PQ[0][1]
######################
## Get LFU or LFU page
######################
def selectEvictPage(self, policy):
r = self.CacheRecency.getFront()
f = self.getHeapMin()
pageToEvit, policyUsed = None, None
if r == f:
pageToEvit, policyUsed = r, -1
elif policy == 0:
pageToEvit, policyUsed = r, 0
elif policy == 1:
pageToEvit, policyUsed = f, 1
# assert pageToEvit in self.CacheRecency
return pageToEvit, policyUsed
def evictPage(self, pg):
assert pg in self.CacheRecency
self.CacheRecency.delete(pg)
############################################
## Choose a page based on the q distribution
############################################
def chooseRandom(self):
r = np.random.rand()
if r < self.W[0]:
return 0
return 1
def addToHistory(self, poly, cacheevict):
histevict = None
if (poly == 0) or (poly == -1 and np.random.rand() < 0.5):
if self.Hist1.size() == self.N:
histevict = self.Hist1.getFront()
assert histevict in self.Hist1
self.Hist1.delete(histevict)
self.Hist1.add(cacheevict)
else:
if self.Hist2.size() == self.N:
histevict = self.Hist2.getFront()
assert histevict in self.Hist2
self.Hist2.delete(histevict)
self.Hist2.add(cacheevict)
if histevict is not None:
del self.evictionTime[histevict]
del self.policyUsed[histevict]
del self.freq[histevict]
def setTime(self, key, t):
if key not in self.TR:
self.TR[key] = 0
self.TR[key] += t
########################################################################################################################################
####REQUEST#############################################################################################################################
########################################################################################################################################
def request(self, page):
starttime = time.time()
page_fault = False
self.time = self.time + 1
###########################
## Clean up
## In case PQ get too large
##########################
if len(self.PQ) > 2 * self.N:
newpq = []
for pg in self.CacheRecency:
newpq.append((self.freq[pg], pg))
heapq.heapify(newpq)
self.PQ = newpq
del newpq
#####################
## Visualization data
#####################
if self.Visualization:
self.X.append(self.time)
self.Y1.append(self.W[0])
self.Y2.append(self.W[1])
##########################
## Process page request
##########################
if page in self.CacheRecency:
st = time.time()
page_fault = False
self.pageHitUpdate(page)
self.setTime('pageHitUpdate', time.time() - st)
else:
#####################################################
## Learning step: If there is a page fault in history
#####################################################
pageevict = None
st = time.time()
reward = np.array([0, 0], dtype=np.float32)
if page in self.Hist1:
pageevict = page
self.Hist1.delete(page)
reward[1] = self.error_discount_rate**(
self.time - self.evictionTime[pageevict])
reward_hat = reward
elif page in self.Hist2:
pageevict = page
self.Hist2.delete(page)
reward[0] = self.error_discount_rate**(
self.time - self.evictionTime[pageevict])
reward_hat = reward
#################
## Update Weights
#################
if pageevict is not None:
self.W = self.W * np.exp(self.learningRate * reward_hat)
self.W = self.W / np.sum(self.W)
# minweight = 0.01
# if self.W[0] < minweight :
# self.W[0] = minweight
# self.W[1] = 1 - self.W[0]
# elif self.W[1] < minweight :
# self.W[1] = minweight
# self.W[0] = 1 - self.W[1]
self.setTime('Hit in history and update weights', time.time() - st)
####################
## Remove from Cache
####################
if self.CacheRecency.size() == self.N:
################
## Choose Policy
################
st = time.time()
act = self.chooseRandom()
self.setTime('chooseRandom', time.time() - st)
st = time.time()
cacheevict, poly = self.selectEvictPage(act)
self.policyUsed[cacheevict] = poly
self.evictionTime[cacheevict] = self.time
self.setTime('selectEvictPage', time.time() - st)
###################
## Remove from Cache and Add to history
###################
st = time.time()
self.evictPage(cacheevict)
self.addToHistory(poly, cacheevict)
self.setTime('selectEvictPage', time.time() - st)
st = time.time()
self.addToCache(page)
self.setTime('addToCache', time.time() - st)
page_fault = True
# st = time.time()
# self.q.put(notInHistory)
# self.sum += notInHistory
# if self.q.qsize() > self.N:
# self.sum -= self.q.get()
# self.NewPages.append(1.0*self.sum / (self.N))
# self.setTime('New pages',time.time()-st)
self.setTime('total', time.time() - starttime)
return page_fault
def get_list_labels(self):
return ['L']
class RecencyAndFrequencyCacheList:
def __init__(self, N, visualization=True):
self.N = N
self.freq = {}
self.PQ = []
self.Cache = CacheLinkedList(N)
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
def __contains__(self, page):
return page in self.Cache
def pageHitUpdate(self, page):
self.cleanPQ()
assert page in self.Cache and page in self.freq
self.Cache.moveBack(page)
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
def addToCache(self, page):
self.cleanPQ()
self.Cache.add(page)
if page not in self.freq:
self.freq[page] = 0
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
def getHeapMin(self):
while self.PQ[0][1] not in self.Cache or self.freq[
self.PQ[0][1]] != self.PQ[0][0]:
heapq.heappop(self.PQ)
return self.PQ[0][1]
######################
## Get LFU or LFU page
######################
def selectEvictPage(self, policy):
self.cleanPQ()
r = self.Cache.getFront()
f = self.getHeapMin()
pageToEvit, policyUsed = None, None
if r == f:
pageToEvit, policyUsed = r, -1
elif policy == 0:
pageToEvit, policyUsed = r, 0
elif policy == 1:
pageToEvit, policyUsed = f, 1
return pageToEvit, policyUsed
def evictPage(self, pg):
assert pg in self.Cache
self.Cache.delete(pg)
def cleanPQ(self):
if len(self.PQ) > 2 * self.N:
newpq = []
for pg in self.Cache:
newpq.append((self.freq[pg], pg))
heapq.heapify(newpq)
self.PQ = newpq
del newpq
def addToHistory(self, poly, cacheevict):
histevict = None
if (poly == 0) or (poly == -1 and np.random.rand() < 0.5):
if self.Hist1.size() == self.N:
histevict = self.Hist1.getFront()
assert histevict in self.Hist1
self.Hist1.delete(histevict)
self.Hist1.add(cacheevict)
else:
if self.Hist2.size() == self.N:
histevict = self.Hist2.getFront()
assert histevict in self.Hist2
self.Hist2.delete(histevict)
self.Hist2.add(cacheevict)
if histevict is not None:
del self.freq[histevict]
return histevict
return None
def deleteHist1(self, page):
self.Hist1.delete(page)
def deleteHist2(self, page):
self.Hist2.delete(page)
def inHistory(self, page):
return page in self.Hist1 or page in self.Hist2
def inCache(self, page):
return page in self.Cache
class LeCaR_new(page_replacement_algorithm):
# def __init__(self, N, visualization = True):
def __init__(self, param):
assert 'cache_size' in param
self.N = int(param['cache_size'])
self.H = int(self.N * int(param['history_size_multiple'])) if 'history_size_multiple' in param else self.N
self.discount_rate = float(param['discount_rate']) if 'discount_rate' in param else 1
self.learning_rate = float(param['learning_rate']) if 'learning_rate' in param else 0
self.initial_weight = float(param['initial_weight']) if 'initial_weight' in param else 0.5
self.Visualization = 'visualize' in param and bool(param['visualize'])
self.discount_rate = 0.005 **(1/self.N)
self.CacheRecency = CacheLinkedList(self.N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(self.H)
self.Hist2 = CacheLinkedList(self.H)
np.random.seed(123)
## Accounting variables
self.time = 0
self.eTime = {}
self.timeIn = {} # Time page was inserted in cache
self.W = np.array([self.initial_weight,1-self.initial_weight], dtype=np.float32)
self.X = []
self.Y1 = []
self.Y2 = []
self.unique = {}
self.unique_cnt = 0
self.pollution_dat_x = []
self.pollution_dat_y = []
self.pollution_dat_y_val = 0
self.pollution_dat_y_sum = []
self.info = {
'lru_misses':0,
'lfu_misses':0,
'lru_count':0,
'lfu_count':0,
}
def get_N(self) :
return self.N
def __contains__(self, q):
return q in self.CacheRecency
def visualize(self, ax_w, ax_h, averaging_window_size):
lbl = []
if self.Visualization:
X = np.array(self.X)
Y1 = np.array(self.Y1)
Y2 = np.array(self.Y2)
ax_w.set_xlim(np.min(X), np.max(X))
ax_h.set_xlim(np.min(X), np.max(X))
ax_w.plot(X,Y1, 'y-', label='W_lru', linewidth=2)
ax_w.plot(X,Y2, 'b-', label='W_lfu', linewidth=1)
#ax_h.plot(self.pollution_dat_x,self.pollution_dat_y, 'g-', label='hoarding',linewidth=3)
#ax_h.plot(self.pollution_dat_x,self.pollution_dat_y, 'k-', linewidth=3)
ax_h.set_ylabel('Hoarding')
ax_w.legend(loc=" upper right")
ax_w.set_title('LeCaR - Adaptive LR')
pollution_sums = self.getPollutions()
temp = np.append(np.zeros(averaging_window_size), pollution_sums[:-averaging_window_size])
pollutionrate = (pollution_sums-temp) / averaging_window_size
ax_h.set_xlim(0, len(pollutionrate))
ax_h.plot(range(len(pollutionrate)), pollutionrate, 'k-', linewidth=3)
# lbl.append(l1)
# lbl.append(l2)
# lbl.append(l3)
return lbl
def getWeights(self):
return np.array([self. X, self.Y1, self.Y2,self.pollution_dat_x,self.pollution_dat_y ]).T
# return np.array([self.pollution_dat_x,self.pollution_dat_y ]).T
def getStats(self):
d={}
d['weights'] = np.array([self. X, self.Y1, self.Y2]).T
d['pollution'] = np.array([self.pollution_dat_x, self.pollution_dat_y ]).T
return d
##############################################################
## There was a page hit to 'page'. Update the data structures
##############################################################
def pageHitUpdate(self, page):
assert page in self.CacheRecency and page in self.freq
self.CacheRecency.moveBack(page)
self.freq[page] += 1
self.timeIn[page] = self.time
heapq.heappush(self.PQ, (self.freq[page], self.timeIn[page], page))
##########################################
## Add a page to cache using policy 'poly'
##########################################
def addToCache(self, page):
assert page not in self.CacheRecency, "Page already in cache"
assert page not in self.timeIn, "Page already in cache"
self.CacheRecency.add(page)
# if page not in self.freq :
# self.freq[page] = 0
self.freq[page] = 1
self.timeIn[page] = self.time
# heapq.heappush(self.PQ, (self.freq[page], page))
heapq.heappush(self.PQ, (self.freq[page], self.timeIn[page], page))
def getHeapMin(self):
# while self.PQ[0][1] not in self.CacheRecency or self.freq[self.PQ[0][1]] != self.PQ[0][0] :
# heapq.heappop(self.PQ)
# return self.PQ[0][1]
#self.print_state(self.PQ.getFreqDic().keys(), self.PQ.getFreqDic())
while self.PQ[0][2] not in self.CacheRecency or self.freq[self.PQ[0][2]] != self.PQ[0][0] or self.timeIn[self.PQ[0][2]] != self.PQ[0][1]:
print "Entering = "
heapq.heappop(self.PQ)
return self.PQ[0][2]
######################
## Get LFU or LFU page
######################
def selectEvictPage(self, policy):
r = self.CacheRecency.getFront()
f = self.getHeapMin()
pageToEvit,policyUsed = None, None
#if r == f :
#pageToEvit,policyUsed = r,-1
if policy == 0:
pageToEvit,policyUsed = r,0
#print "LRU ------------------------------"
elif policy == 1:
pageToEvit,policyUsed = f,1
return pageToEvit,policyUsed
def evictPage(self, pg):
assert pg in self.CacheRecency
assert pg in self.timeIn
self.CacheRecency.delete(pg)
del self.timeIn[pg]
def getQ(self):
lamb = 0.05
return (1-lamb)*self.W + lamb
############################################
## Choose a page based on the q distribution
############################################
def chooseRandom(self):
r = np.random.rand()
if r < self.W[0] :
return 0
return 1
def addToHistory(self, poly, cacheevict):
histevict = None
if (poly == 0) or (poly==-1 and np.random.rand() <0.5):
if self.Hist1.size() == self.H :
histevict = self.Hist1.getFront()
assert histevict in self.Hist1
self.Hist1.delete(histevict)
self.Hist1.add(cacheevict)
self.info['lru_count'] += 1
else:
if self.Hist2.size() == self.H :
histevict = self.Hist2.getFront()
assert histevict in self.Hist2
self.Hist2.delete(histevict)
self.Hist2.add(cacheevict)
self.info['lfu_count'] += 1
if histevict is not None :
del self.freq[histevict]
del self.eTime[histevict]
########################################################################################################################################
####REQUEST#############################################################################################################################
########################################################################################################################################
def request(self,page) :
page_fault = False
self.time = self.time + 1
#print "request = ", page
#self.print_state(self.CacheRecency, self.freq)
#print "PQ = ", self.PQ
#print "heapq= ", heapq
#self.print_state(self.PQ.getFreqDic().keys(), self.PQ.getFreqDic())
###########################
## Clean up
## In case PQ get too large
##########################
if len(self.PQ) > self.N:
#print "entering cleaning"
newpq = []
for pg in self.CacheRecency:
newpq.append((self.freq[pg], self.timeIn[pg], pg))
heapq.heapify(newpq)
self.PQ = newpq
del newpq
#print "PQ = ", self.PQ
#####################
## Visualization data
#####################
if self.Visualization:
self.X.append(self.time)
self.Y1.append(self.W[0])
self.Y2.append(self.W[1])
##########################
## Process page request
##########################
if page in self.CacheRecency:
page_fault = False
self.pageHitUpdate(page)
else :
#####################################################
## Learning step: If there is a page fault in history
#####################################################
pageevict = None
reward = np.array([0,0], dtype=np.float32)
if page in self.Hist1:
pageevict = page
self.Hist1.delete(page)
#reward[0] = -1
reward[0] = -self.discount_rate **(self.time-self.eTime[page])
self.info['lru_misses'] +=1
elif page in self.Hist2:
pageevict = page
self.Hist2.delete(page)
#reward[1] = -1
reward[1] = -self.discount_rate **(self.time-self.eTime[page])
self.info['lfu_misses'] +=1
#################
## Update Weights
#################
if pageevict is not None :
self.W = self.W * np.exp(self.learning_rate * reward)
self.W = self.W / np.sum(self.W)
####################
## Remove from Cache
####################
if self.CacheRecency.size() == self.N:
################
## Choose Policy
################
act = self.chooseRandom()
cacheevict,poly = self.selectEvictPage(act)
###################
## Remove from Cache and Add to history
###################
self.eTime[cacheevict] = self.time
self.evictPage(cacheevict)
self.addToHistory(poly, cacheevict)
#print "evict = ", cacheevict
self.addToCache(page)
page_fault = True
## Count pollution
if page_fault:
self.unique_cnt += 1
self.unique[page] = self.unique_cnt
if self.time % self.N == 0:
pollution = 0
for pg in self.CacheRecency:
if self.unique_cnt - self.unique[pg] >= 2*self.N:
pollution += 1
self.pollution_dat_x.append(self.time)
self.pollution_dat_y.append(100* pollution / self.N)
self.pollution_dat_y_val += 100* self.pollution / self.N
self.pollution_dat_y_sum.append(self.pollution_dat_y_val)
assert self.CacheRecency.size() <= self.N
return page_fault
def getPollutions(self):
return self.pollution_dat_y_sum
def getLearningRates(self):
return self.learning_rates
def get_list_labels(self) :
return ['L']
def __init__(self, param):
assert 'cache_size' in param
# assert 'history_size_multiple' in param
self.N = int(param['cache_size'])
self.H = int(self.N * int(param['history_size_multiple'])) if 'history_size_multiple' in param else self.N
self.learning_rate = float(param['learning_rate']) if 'learning_rate' in param else 0
self.discount_rate = float(param['discount_rate']) if 'discount_rate' in param else 1
# self.discount_rate = (float(param['discount_rate']) if 'discount_rate' in param else 0) ** (1/self.N)
# self.discount_rate = 0.05**(1/self.N)
self.Visualization = 'visualize' in param and bool(param['visualize'])
self.lamb = 0.05
self.learning_rate = 0.1
self.learning_rate_lfu= 0.1
self.learning_rate_lru= 0.1
self.CacheRecency = CacheLinkedList(self.N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(self.H)
self.Hist2 = CacheLinkedList(self.H)
np.random.seed(123)
self.PageCount = 0
self.CacheHit = 0
self.PreviousHR = 0.0
self.NewHR = 0.0
self.PreviousChangeInHR = 0.0
self.NewChangeInHR =0.0
self.PreviousLR= 0.45
self.NewLR =0.45
self.CacheHitList = []
self.counter = 0
self.learning_rates = []
self.SampleHR =[]
self.SAMPLE_SIZE = 20 * self.N
self.SampleHitQ = queue.Queue(maxsize= self.SAMPLE_SIZE)
self.SampleCacheHit = 0
## Accounting variables
self.time = 0
self.W = np.array([.5,.5], dtype=np.float32)
self.PreviousW = np.array([.5,.5], dtype=np.float32)
self.NewW = np.array([.5,.5], dtype=np.float32)
self.qUsed = {}
self.eTime = {}
self.X = []
self.Y1 = []
self.Y2 = []
self.unique = {}
self.unique_cnt = 0
self.pollution_dat_x = []
self.pollution_dat_y = []
self.pollution_dat_y_val = 0
self.pollution_dat_y_sum = []
self.pollution =0
class LeCaR_q(page_replacement_algorithm):
def __init__(self, N, visualization=True):
self.N = N
self.CacheRecency = CacheLinkedList(N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
## Config variables
self.error_discount_rate = (0.005)**(1.0 / N)
self.learning_rate = 0.5
##
self.evictionTime = {}
## Accounting variables
self.time = 0
self.W = np.array([.5, .5], dtype=np.float32)
self.Visualization = visualization
self.X = []
self.Y1 = []
self.Y2 = []
self.gamma = 0.05 # uniform distribution mixture parameter
self.q_used = {}
self.unique = {}
self.unique_cnt = 0
self.pollution_dat_x = []
self.pollution_dat_y = []
def get_N(self):
return self.N
def visualize(self, plt):
lbl = []
if self.Visualization:
X = np.array(self.X)
Y1 = np.array(self.Y1)
Y2 = np.array(self.Y2)
ax = plt.subplot(2, 1, 1)
ax.set_xlim(np.min(X), np.max(X))
l3, = plt.plot(self.pollution_dat_x,
self.pollution_dat_y,
'g-',
label='hoarding',
linewidth=3)
l1, = plt.plot(X, Y1, 'y-', label='W_lru', linewidth=2)
l2, = plt.plot(X, Y2, 'b-', label='W_lfu', linewidth=1)
lbl.append(l1)
lbl.append(l2)
lbl.append(l3)
return lbl
def getWeights(self):
return np.array([
self.X, self.Y1, self.Y2, self.pollution_dat_x,
self.pollution_dat_y
]).T
# return np.array([self.pollution_dat_x,self.pollution_dat_y ]).T
def getStats(self):
d = {}
d['weights'] = np.array([self.X, self.Y1, self.Y2]).T
d['pollution'] = np.array([self.pollution_dat_x,
self.pollution_dat_y]).T
return d
##############################################################
## There was a page hit to 'page'. Update the data structures
##############################################################
def pageHitUpdate(self, page):
assert page in self.CacheRecency and page in self.freq
self.CacheRecency.moveBack(page)
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
##########################################
## Add a page to cache using policy 'poly'
##########################################
def addToCache(self, page):
self.CacheRecency.add(page)
if page not in self.freq:
self.freq[page] = 0
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
def getHeapMin(self):
while self.PQ[0][1] not in self.CacheRecency or self.freq[
self.PQ[0][1]] != self.PQ[0][0]:
heapq.heappop(self.PQ)
return self.PQ[0][1]
######################
## Get LFU or LFU page
######################
def selectEvictPage(self, policy):
r = self.CacheRecency.getFront()
f = self.getHeapMin()
pageToEvit, policyUsed = None, None
if r == f:
pageToEvit, policyUsed = r, -1
elif policy == 0:
pageToEvit, policyUsed = r, 0
elif policy == 1:
pageToEvit, policyUsed = f, 1
return pageToEvit, policyUsed
def evictPage(self, pg):
assert pg in self.CacheRecency
self.CacheRecency.delete(pg)
############################################
## Choose a page based on the q distribution
############################################
def chooseRandom(self, q):
r = np.random.rand()
if r < q[0]:
return 0
return 1
def addToHistory(self, poly, cacheevict):
histevict = None
if (poly == 0) or (poly == -1 and np.random.rand() < 0.5):
if self.Hist1.size() == self.N:
histevict = self.Hist1.getFront()
assert histevict in self.Hist1
self.Hist1.delete(histevict)
self.Hist1.add(cacheevict)
else:
if self.Hist2.size() == self.N:
histevict = self.Hist2.getFront()
assert histevict in self.Hist2
self.Hist2.delete(histevict)
self.Hist2.add(cacheevict)
if histevict is not None:
del self.evictionTime[histevict]
del self.freq[histevict]
del self.q_used[histevict]
########################################################################################################################################
####REQUEST#############################################################################################################################
########################################################################################################################################
def request(self, page):
page_fault = False
self.time = self.time + 1
###########################
## Clean up
## In case PQ get too large
##########################
if len(self.PQ) > 2 * self.N:
newpq = []
for pg in self.CacheRecency:
newpq.append((self.freq[pg], pg))
heapq.heapify(newpq)
self.PQ = newpq
del newpq
#####################
## Visualization data
#####################
if self.Visualization:
self.X.append(self.time)
self.Y1.append(self.W[0])
self.Y2.append(self.W[1])
##########################
## Process page request
##########################
if page in self.CacheRecency:
page_fault = False
self.pageHitUpdate(page)
else:
#####################################################
## Learning step: If there is a page fault in history
#####################################################
pageevict = None
reward = np.array([0, 0], dtype=np.float32)
if page in self.Hist1:
pageevict = page
self.Hist1.delete(page)
reward[1] = self.error_discount_rate**(
self.time -
self.evictionTime[pageevict]) / self.q_used[pageevict][0]
elif page in self.Hist2:
pageevict = page
self.Hist2.delete(page)
reward[0] = self.error_discount_rate**(
self.time -
self.evictionTime[pageevict]) / self.q_used[pageevict][1]
#################
## Update Weights
#################
if pageevict is not None:
self.W = self.W * np.exp(self.learning_rate * reward)
self.W = self.W / np.sum(self.W)
####################
## Remove from Cache
####################
if self.CacheRecency.size() == self.N:
################
## Choose Policy
################
q = (1 - self.gamma) * self.W + self.gamma / 2
act = self.chooseRandom(q)
cacheevict, poly = self.selectEvictPage(act)
self.evictionTime[cacheevict] = self.time
self.q_used[cacheevict] = q
###################
## Remove from Cache and Add to history
###################
self.evictPage(cacheevict)
self.addToHistory(poly, cacheevict)
self.addToCache(page)
page_fault = True
## Count pollution
if page_fault:
self.unique_cnt += 1
self.unique[page] = self.unique_cnt
if self.time % self.N == 0:
pollution = 0
for pg in self.CacheRecency:
if self.unique_cnt - self.unique[pg] >= 2 * self.N:
pollution += 1
self.pollution_dat_x.append(self.time)
self.pollution_dat_y.append(100 * pollution / self.N)
return page_fault
def get_list_labels(self):
return ['L']
class LOMP(page_replacement_algorithm):
def __init__(self, N):
self.N = N
self.CacheRecency = CacheLinkedList(N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
## Config variables
self.epsilon = 0.05
self.error_discount_rate = (0.005)**(1.0 / N)
self.policy_space_size = 4
self.Gamma = 0.5
self.minWeight = 0.01
##
self.evictionTime = {}
self.policyUsed = {}
self.weightsUsed = {}
self.qUsed = {}
self.freq = {}
## Accounting variables
self.time = 0
self.unif = self.Gamma * np.ones(
self.policy_space_size, dtype=np.float64) / self.policy_space_size
self.W = np.ones(self.policy_space_size,
dtype=np.float64) / self.policy_space_size
self.X = np.array([], dtype=np.int32)
self.Y = np.array([])
###
self.q = Queue.Queue()
self.sum = 0
self.NewPages = []
def get_N(self):
return self.N
def visualize(self, plt):
# print(np.min(self.X), np.max(self.X))
ax = plt.subplot(2, 1, 1)
ax.set_xlim(np.min(self.X), np.max(self.X))
lbl = []
for i in range(0, self.policy_space_size):
l, = plt.plot(self.X, self.Y[:, i], label='W_%d' % i)
lbl.append(l)
l3, = plt.plot(self.X,
self.NewPages,
'g-',
label='New Pages',
alpha=0.6)
lbl.append(l3)
return lbl
############################################
## Choose a page based on the q distribution
############################################
def chooseRandom(self, w):
tmp = 1.0 * w / np.sum(w, dtype=np.float64)
cdf = np.cumsum(tmp, dtype=np.float64)
r = np.random.rand()
for policy, pr in enumerate(cdf):
if r < pr:
return policy
# print np.sum(w), w / np.sum(w), cdf
return len(w) - 1
##############################################################
## There was a page hit to 'page'. Update the data structures
##############################################################
def pageHitUpdate(self, page):
assert page in self.CacheRecency and page in self.freq
self.CacheRecency.moveBack(page)
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
##########################################
## Add a page to cache using policy 'poly'
##########################################
def addToCache(self, page):
self.CacheRecency.add(page)
if page not in self.freq:
self.freq[page] = 0
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
def getHeapMin(self):
if len(self.PQ) < self.N:
print self.PQ
assert len(self.PQ) >= self.N, 'PQ should be full %d' % len(self.PQ)
while self.PQ[0][1] not in self.CacheRecency or self.freq[
self.PQ[0][1]] != self.PQ[0][0]:
heapq.heappop(self.PQ)
return self.PQ[0][1]
######################
## Get LFU or LFU page
######################
def selectEvictPage(self, policy):
if np.random.rand() < 1.0 * policy / (self.policy_space_size - 1):
return self.CacheRecency.getFront()
else:
return self.getHeapMin()
def evictPage(self, pg):
assert pg in self.CacheRecency
self.CacheRecency.delete(pg)
def addToHistory(self, cacheevict):
histevict = None
if self.Hist1.size() == self.N:
histevict = self.Hist1.getFront()
assert histevict in self.Hist1
self.Hist1.delete(histevict)
self.Hist1.add(cacheevict)
if histevict is not None:
del self.evictionTime[histevict]
del self.policyUsed[histevict]
del self.freq[histevict]
########################################################################################################################################
####REQUEST#############################################################################################################################
########################################################################################################################################
def request(self, page):
page_fault = False
self.time = self.time + 1
# if self.time % self.learning_phase == 0 :
# self.learning = not self.learning
###########################
## Clean up
## In case PQ get too large
##########################
if len(self.PQ) > 2 * self.N:
newpq = []
for pg in self.CacheRecency:
newpq.append((self.freq[pg], pg))
heapq.heapify(newpq)
self.PQ = newpq
del newpq
#####################
## Visualization data
#####################
if self.time == 1:
self.Y = np.append(self.Y, self.W)
else:
self.Y = np.vstack((self.Y, self.W))
self.X = np.append(self.X, self.time)
notInHistory = 0
##########################
## Process page request
##########################
if page in self.CacheRecency:
page_fault = False
self.pageHitUpdate(page)
else:
#####################################################
## Learning step: If there is a page fault in history
#####################################################
if page in self.Hist1:
self.Hist1.delete(page)
et = self.evictionTime[page]
pu = self.policyUsed[page]
qu = self.qUsed[page]
cost = np.zeros(self.policy_space_size, dtype=np.float64)
cost[pu] = self.error_discount_rate**(self.time - et)
cost_hat = cost / qu
#################
## Update Weights
#################
self.W = self.W * (1.0 - self.epsilon * cost_hat)
cost[pu] = np.min(cost[pu], self.minWeight)
self.W = self.W / np.sum(self.W)
# print np.sum(self.W)
assert np.sum(self.W) > 0.00000001, 'ERROR: W is zero'
else:
notInHistory = 1
####################
## Remove from Cache
####################
if self.CacheRecency.size() == self.N:
################
## Choose Policy
################
q = (1.0 - self.Gamma) * self.W + self.unif
act = self.chooseRandom(q)
cacheevict = self.selectEvictPage(act)
self.policyUsed[cacheevict] = act
self.evictionTime[cacheevict] = self.time
self.qUsed[cacheevict] = q
###################
## Remove from Cache and Add to history
###################
self.evictPage(cacheevict)
self.addToHistory(cacheevict)
self.addToCache(page)
page_fault = True
self.q.put(notInHistory)
self.sum += notInHistory
if self.q.qsize() > self.N:
self.sum -= self.q.get()
self.NewPages.append(1.0 * self.sum / self.N)
return page_fault
def get_list_labels(self):
return ['L']
class OLCR(page_replacement_algorithm):
def __init__(self, N):
self.N = N
self.CacheRecency = CacheLinkedList(N)
self.freq = {}
self.PQ = []
self.Hist1 = CacheLinkedList(N)
self.Hist2 = CacheLinkedList(N)
## Config variables
self.error_discount_rate = (0.005)**(1.0 / N)
##
self.policy = 0
self.evictionTime = {}
self.policyUsed = {}
self.pUsed = {}
self.param = {}
## Accounting variables
self.time = 0
###
self.q = Queue.Queue()
self.sum = 0
self.NewPages = []
self.c_hits = 0
self.h_miss = 0
self.learning = True
input_units = 2 * self.N
hidden_units = 3
hidden_units2 = 2
output_units = 2
self.X = tf.placeholder(dtype=tf.int32, shape=[None, 2])
self.P = tf.placeholder(dtype=tf.float32, shape=[None, 1])
self.C_r = tf.placeholder(dtype=tf.float32, shape=[None])
self.C_f = tf.placeholder(dtype=tf.float32, shape=[None])
self.W = tf.Variable(tf.random_uniform([input_units, hidden_units]))
self.b = tf.Variable(tf.ones([hidden_units]))
self.W_hidden = tf.Variable(
tf.random_uniform([hidden_units, hidden_units2]))
self.b_hidden = tf.Variable(tf.ones([hidden_units2]))
self.W_hidden2 = tf.Variable(
tf.random_uniform([hidden_units2, output_units]))
self.b_hidden2 = tf.Variable(tf.ones([output_units]))
# self.predict = tf.sigmoid(tf.matmul(self.X, self.W))
# self.predict = tf.sigmoid(tf.slice(self.W, self.X[0,0],[1]) + tf.slice(self.W, self.X[0,1]+self.N,[1]))
idx1 = tf.slice(self.X, [0, 0], [-1, 1])
idx2 = tf.slice(self.X, [0, 1], [-1, 1]) + self.N
## Neuron 1
# self.w1 = tf.slice(self.W, [0, idx1[0]],[1,1])
# self.w2 = tf.slice(self.W, [0, idx2[0]+self.N,[1])
self.w11 = tf.slice(self.W[:, 0], idx1[0], [1])
self.w22 = tf.slice(self.W[:, 1], idx2[0], [1])
self.w13 = tf.slice(self.W[:, 2], idx1[0], [1])
self.w23 = tf.slice(self.W[:, 2], idx2[0], [1])
## Hidden layer
self.neuron1 = tf.sigmoid(self.w11 + self.b[0]) # only x1
self.neuron2 = tf.sigmoid(self.w22 + self.b[1]) # only x2
self.neuron3 = tf.sigmoid(self.w13 + self.w23 + self.b[2]) # x1+x2
temp = tf.concat([[self.neuron1], [self.neuron2], [self.neuron3]], 1)
self.logit1 = tf.contrib.layers.flatten(temp) # 1 x hidden_units
self.logit2 = tf.sigmoid(tf.matmul(self.logit1, self.W_hidden))
self.logit3 = tf.sigmoid(tf.matmul(self.logit2, self.W_hidden2))
# self.predict = tf.nn.softmax(self.logit2)
self.predict = tf.nn.softmax(self.logit3)
# self.cost_r = -tf.reduce_mean(self.C_r * tf.log(1-self.predict[0,0]) + (1-self.C_r) * tf.log(self.predict[0,0]))
# self.cost_f = -tf.reduce_mean(self.C_f * tf.log(1-self.predict[0,1]) + (1-self.C_f) * tf.log(self.predict[0,1]))
# self.cost = self.cost_r + self.cost_f
self.cost = -tf.reduce_mean(self.C_r * tf.log(1 - self.predict[0, 0]) +
self.C_f * tf.log(1 - self.predict[0, 1]))
learning_rate = 0.01
self.optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(self.cost)
##################################
self.X_holder = []
self.P_holder = []
self.R_holder = []
self.F_holder = []
self.train_batch_size = 4 * self.N
init = tf.global_variables_initializer()
self.sess = tf.Session()
self.sess.run(init)
def get_N(self):
return self.N
def visualize(self, plt):
return []
##############################################################
## There was a page hit to 'page'. Update the data structures
##############################################################
def pageHitUpdate(self, page):
assert page in self.CacheRecency and page in self.freq
self.CacheRecency.moveBack(page)
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
##########################################
## Add a page to cache using policy 'poly'
##########################################
def addToCache(self, page):
self.CacheRecency.add(page)
if page not in self.freq:
self.freq[page] = 0
self.freq[page] += 1
heapq.heappush(self.PQ, (self.freq[page], page))
def getHeapMin(self):
assert len(self.PQ) >= self.N, 'PQ should be full %d' % len(self.PQ)
while self.PQ[0][1] not in self.CacheRecency or self.freq[
self.PQ[0][1]] != self.PQ[0][0]:
heapq.heappop(self.PQ)
return self.PQ[0][1]
######################
## Get LFU or LFU page
## return page, poly
######################
def selectEvictPage(self, P):
assert P >= 0 and P <= 1.1, 'P = %f' % P
if np.random.rand() < P:
return self.CacheRecency.getFront(), 0
else:
return self.getHeapMin(), 1
def evictPage(self, pg):
assert pg in self.CacheRecency
self.CacheRecency.delete(pg)
############################################
## Choose a page based on the q distribution
############################################
def chooseRandom(self):
r = np.random.rand()
if r < self.W[0]:
return 0
return 1
def addToHistory(self, poly, cacheevict):
histevict = None
if poly == 0:
if self.Hist1.size() == self.N:
histevict = self.Hist1.getFront()
assert histevict in self.Hist1
self.Hist1.delete(histevict)
self.Hist1.add(cacheevict)
elif poly == 1:
if self.Hist2.size() == self.N:
histevict = self.Hist2.getFront()
assert histevict in self.Hist2
self.Hist2.delete(histevict)
self.Hist2.add(cacheevict)
if histevict is not None:
del self.evictionTime[histevict]
del self.policyUsed[histevict]
del self.freq[histevict]
del self.pUsed[histevict]
del self.param[histevict]
########################################################################################################################################
####REQUEST#############################################################################################################################
########################################################################################################################################
def request(self, page):
page_fault = False
self.time = self.time + 1
#######################
## Train
#######################
if len(self.X_holder) >= self.train_batch_size:
X_1 = np.array(self.X_holder)
R = np.array(self.R_holder)
F = np.array(self.F_holder)
# t1 = tf.one_hot(X_1[:,0], depth=self.N)
# t2 = tf.one_hot(X_1[:,1], depth=self.N)
# X_2 = tf.concat([t1,t2], 1)
cost_bef = self.sess.run(self.cost,
feed_dict={
self.X: X_1,
self.C_r: R,
self.C_f: F
})
for _ in range(0, 50):
self.sess.run(self.optimizer,
feed_dict={
self.X: X_1,
self.C_r: R,
self.C_f: F
})
cost_after = self.sess.run(self.cost,
feed_dict={
self.X: X_1,
self.C_r: R,
self.C_f: F
})
# print 'bef:%f - after:%f = %f' % (cost_bef, cost_after,cost_bef-cost_after)
self.X_holder = []
self.P_holder = []
self.R_holder = []
self.F_holder = []
###########################
## Clean up
## In case PQ get too large
##########################
if len(self.PQ) > 2 * self.N:
newpq = []
for pg in self.CacheRecency:
newpq.append((self.freq[pg], pg))
heapq.heapify(newpq)
self.PQ = newpq
del newpq
page_outcome = -1
##########################
## Process page request
##########################
if page in self.CacheRecency:
page_fault = False
self.pageHitUpdate(page)
page_outcome = -1
else:
#####################################################
## Learning step: If there is a page fault in history
#####################################################
if page in self.Hist1 or page in self.Hist2:
page_outcome = 1
X = self.param[page]
P = self.pUsed[page]
e = self.error_discount_rate**(self.time -
self.evictionTime[page])
## TODO Consider dividing e by P
self.X_holder.append(X)
self.P_holder.append(P)
if page in self.Hist1:
self.Hist1.delete(page)
self.R_holder.append(1)
self.F_holder.append(0)
else:
self.Hist2.delete(page)
self.R_holder.append(0)
self.F_holder.append(1)
else:
page_outcome = 2
####################
## Remove from Cache
####################
if self.CacheRecency.size() == self.N:
################
## Choose Policy
################
# t1 = tf.one_hot([self.c_hits], depth=self.N)
# t2 = tf.one_hot([self.h_miss], depth=self.N)
# X_2 = tf.concat([t1,t2], 1)
if np.random.rand() < 0.5:
P = self.sess.run(self.predict,
feed_dict={
self.X:
np.array([[self.c_hits,
self.h_miss]])
})[0, 0]
else:
P = np.random.rand()
# print self.sess.run(self.unnormweight,feed_dict={self.X:np.array([[self.c_hits, self.h_miss]])})
# print 'P = ', P
# assert np.sum(P) <= 1.0, np.sum(P)
# P = np.random.rand()
cacheevict, poly = self.selectEvictPage(P)
self.policyUsed[cacheevict] = poly
self.evictionTime[cacheevict] = self.time
self.pUsed[cacheevict] = P
self.param[cacheevict] = [self.c_hits, self.h_miss]
###################
## Remove from Cache and Add to history
###################
self.evictPage(cacheevict)
self.addToHistory(poly, cacheevict)
self.addToCache(page)
page_fault = True
self.q.put(page_outcome)
if page_outcome == 1:
self.c_hits += 1
elif page_outcome == 2:
self.h_miss += 1
if self.q.qsize() >= self.N:
temp = self.q.get()
if temp == 1:
self.c_hits -= 1
elif temp == 2:
self.h_miss -= 1
assert self.c_hits >= 0 and self.c_hits < self.N
assert self.h_miss >= 0 and self.h_miss < self.N
return page_fault
def get_list_labels(self):
return ['L']