class LaCReME(page_replacement_algorithm): def __init__(self, N): self.N = N self.CacheRecency = Disk(N) self.CacheFrequecy = priorityqueue(N) self.Hist1 = Disk(N) self.Hist2 = priorityqueue(N) ## Config variables self.decayRate = 1 self.epsilon = 0.05 self.lamb = 0.05 self.learning_phase = N / 2 # self.error_discount_rate = (0.005)**(1.0/N) ## TODO ADD BACK self.error_discount_rate = 1 ## self.learning = True self.policy = 0 self.evictionTime = {} self.policyUsed = {} self.weightsUsed = {} self.freq = {} ## TODO add decay_time and decay_factor self.decay_time = N self.decay_factor = 1 ## 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([]) 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)) l1, = plt.plot(self.X, self.Y1, 'b-', label='W_lru') l2, = plt.plot(self.X, self.Y2, 'r-', label='W_lfu') return [l1, l2] ########################################## ## Add a page to cache using policy 'poly' ########################################## def addToCache(self, page, pagefreq=0): self.CacheRecency.add(page) self.CacheFrequecy.add(page) self.CacheRecency.increaseCount(page, amount=pagefreq) self.CacheFrequecy.increase(page, amount=pagefreq) ###################### ## Get LFU or LFU page ###################### def selectEvictPage(self, policy): r = self.CacheRecency.getIthPage(0) f = self.CacheFrequecy.peaktop() 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): self.CacheRecency.delete(pg) self.CacheFrequecy.delete(pg) ############################################################## ## There was a page hit to 'page'. Update the data structures ############################################################## def pageHitUpdate(self, page): if page in self.CacheRecency and page in self.CacheFrequecy: self.CacheRecency.moveBack(page) self.CacheRecency.increaseCount(page) self.CacheFrequecy.increase(page) ######################### ## Get the Q distribution ######################### def getQ(self): return (1 - self.lamb) * self.W + self.lamb * np.ones(2) / 2 ############################################ ## Choose a page based on the q distribution ############################################ def chooseRandom(self): q = self.getQ() r = np.random.rand() for i, p in enumerate(q): if r < p: return i return len(q) - 1 def updateWeight(self, cost): self.W = self.W * (1 - self.epsilon * cost) self.W = self.W / np.sum(self.W) ######################################################################################################################################## ####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 ##################### ## Visualization data ##################### prob = self.getQ() self.X = np.append(self.X, self.time) self.Y1 = np.append(self.Y1, prob[0]) self.Y2 = np.append(self.Y2, prob[1]) if self.time % self.N == 0: self.CacheFrequecy.decay(self.decay_factor) self.Hist2.decay(self.decay_factor) ########################## ## Process page request ########################## if page in self.CacheFrequecy: page_fault = False self.pageHitUpdate(page) else: ##################################################### ## Learning step: If there is a page fault in history ##################################################### pageevict, histpage_freq = None, 1 policyUsed = -1 if page in self.Hist1: pageevict = page histpage_freq = self.Hist1.getCount(page) self.Hist1.delete(page) policyUsed = 0 self.W[0] = self.W[0] * ( 1 - self.epsilon) if self.policyUsed[page] != -1 else self.W[0] elif page in self.Hist2: pageevict = page histpage_freq = self.Hist2.getCount( page) ## Get the page frequency in history self.Hist2.delete(page) policyUsed = 1 self.W[1] = self.W[1] * ( 1 - self.epsilon) if self.policyUsed[page] != -1 else self.W[1] self.W = self.W / np.sum(self.W) # if pageevict is not None : # q = self.weightsUsed[pageevict] # # err = self.error_discount_rate ** (self.time - self.evictionTime[pageevict]) # err = 1 # reward = np.array([0,0], dtype=np.float32) # if policyUsed == 0 : # LRU # reward[0] = err # if policyUsed == 1: # reward[1] = err # reward_hat = reward ################# ## Update Weights ################# # if self.policyUsed[pageevict] != -1 : # # self.W = self.W * np.exp(self.lamb * reward_hat / 2) # self.W = self.W * (1 - reward*self.epsilon) # self.W = self.W / np.sum(self.W) #################### ## Remove from Cache #################### if self.CacheRecency.size() == self.N: ################ ## Choose Policy ################ # if not self.learning : # act = np.argmax(self.getQ()) # else : # act = self.chooseRandom() act = np.argmax(self.W) ## REMOVE cacheevict, poly = self.selectEvictPage(act) pagefreq = self.CacheFrequecy.getCount(cacheevict) self.policyUsed[cacheevict] = poly self.weightsUsed[cacheevict] = self.getQ() self.evictionTime[cacheevict] = self.time ## TODO ADD BACK # if not self.learning : # self.policyUsed[cacheevict] = -1 ################### ## Evict to history ################### histevict = None if (poly == 0) or (poly == -1 and np.random.rand() < 0.5): if self.Hist1.size() == self.N: histevict = self.Hist1.getIthPage(0) self.Hist1.delete(histevict) self.Hist1.add(cacheevict) self.Hist1.setCount(cacheevict, pagefreq) else: if self.Hist2.size() == self.N: histevict = self.Hist2.popmin() self.Hist2.add(cacheevict) self.Hist2.increase(cacheevict, pagefreq - 1) if histevict is not None: del self.evictionTime[histevict] del self.policyUsed[histevict] del self.weightsUsed[histevict] self.evictPage(cacheevict) self.addToCache(page, pagefreq=histpage_freq) page_fault = True return page_fault def get_list_labels(self): return ['L']
class LaCReME_T1T2(page_replacement_algorithm): def __init__(self, N): self.N = N self.T1 = Disk(N) self.T2 = Disk(N) self.Hist1 = Disk(N) self.Hist2 = Disk(N) ## Config variables self.epsilon = 0.90 self.lamb = 0.05 self.error_discount_rate = (0.005)**(1.0 / N) # self.learning_phase = N/2 # self.error_discount_rate = 1 ## self.policy = 0 self.evictionTime = {} self.policyUsed = {} self.weightsUsed = {} self.freq = {} ## TODO add decay_time and decay_factor self.decay_time = N self.decay_factor = 1 ## 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 = [] 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)) l1, = plt.plot(self.X, self.Y1, 'b-', label='W_lru') l2, = plt.plot(self.X, self.Y2, 'r-', label='W_lfu') l3, = plt.plot(self.X, self.NewPages, 'g-', label='New Pages', alpha=0.6) return [l1, l2, l3] ############################################################## ## There was a page hit to 'page'. Update the data structures ############################################################## def pageHitUpdate(self, page): if page in self.T1: self.T1.delete(page) self.T2.add(page) else: self.T2.moveBack(page) ######################### ## Get the Q distribution ######################### def getQ(self): return (1 - self.lamb) * self.W + self.lamb * np.ones(2) / 2 ############################################ ## Choose a page based on the q distribution ############################################ def chooseRandom(self): q = self.getQ() r = np.random.rand() for i, p in enumerate(q): if r < p: return i return len(q) - 1 def updateWeight(self, cost): self.W = self.W * (1 - self.epsilon * cost) self.W = self.W / np.sum(self.W) ######################################################################################################################################## ####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 ##################### ## Visualization data ##################### prob = self.getQ() self.X = np.append(self.X, self.time) self.Y1 = np.append(self.Y1, prob[0]) self.Y2 = np.append(self.Y2, prob[1]) notInHistory = 0 ########################## ## Process page request ########################## t1 = self.T1.size() t2 = self.T2.size() assert t1 + t2 <= self.N if page in self.T1 or page in self.T2: page_fault = False self.pageHitUpdate(page) else: ##################################################### ## Learning step: If there is a page fault in history ##################################################### pageevict = None inHist = False policyUsed = -1 if page in self.Hist1: pageevict = page self.Hist1.delete(page) policyUsed = 0 inHist = True elif page in self.Hist2: pageevict = page self.Hist2.delete(page) policyUsed = 1 inHist = True else: notInHistory = 1 if pageevict is not None: q = self.weightsUsed[pageevict] # err = self.error_discount_rate ** (self.time - self.evictionTime[pageevict]) err = 1 reward = np.array([0, 0], dtype=np.float32) if policyUsed == 0: # LRU reward[1] = err if policyUsed == 1: reward[0] = err reward_hat = reward / q ################# ## Update Weights ################# if self.policyUsed[pageevict] != -1: self.W = self.W * np.exp(self.lamb * reward_hat / 2) self.W = self.W / np.sum(self.W) #################### ## Remove from Cache #################### if t1 + t2 == self.N: ################ ## Choose Policy ################ act = self.chooseRandom() if t1 == self.N or (act == 0 and t1 > 0): cacheevict = self.T1.popFront() else: cacheevict = self.T2.popFront() self.policyUsed[cacheevict] = act self.weightsUsed[cacheevict] = self.getQ() self.evictionTime[cacheevict] = self.time ################### ## Evict to history ################### histevict = None if act == 0: if self.Hist1.size() == self.N: histevict = self.Hist1.getFront() self.Hist1.delete(histevict) self.Hist1.add(cacheevict) else: if self.Hist2.size() == self.N: histevict = self.Hist2.getFront() self.Hist2.delete(histevict) self.Hist2.add(cacheevict) if histevict is not None: del self.evictionTime[histevict] del self.policyUsed[histevict] del self.weightsUsed[histevict] if inHist: self.T2.add(page) else: self.T1.add(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 ExpertLearning_v2(page_replacement_algorithm): def __init__(self, N): self.T = [] self.N = N self.disk = Disk(N) self.freq = {} ## Training variables self.X, self.Y = [], [] self.reward = [] self.regret = [] ## Config variables self.batchsize = N self.numbatch = 5 self.discountrate = 0.9 self.error = 0.5 self.reduceErrorRate = 0.975 ## Aux variables self.cachebuff = dequecustom() self.Xbuff = dequecustom() self.Ybuff = dequecustom() self.pageHitBuff = dequecustom() self.hist = dequecustom() self.batchsizeBuff = dequecustom() ## Accounting variables self.currentPageHits = 0 self.current = 0 self.uniquePages = Counter() ## Batch action variable self.action = [0] #self.discount = 0.9 #self.sampleCount = 0 #self.trainingSampleSize = 5 * N ## start tf tf.reset_default_graph() self.input = tf.placeholder(shape=[1, self.N], dtype=tf.float32) W1 = tf.Variable(tf.random_uniform([self.N, 8], 0, 0.01)) out1 = tf.sigmoid(tf.matmul(self.input, W1)) W2 = tf.Variable(tf.random_uniform([8, 2], 0, 0.01)) self.out = tf.matmul(out1, W2) self.predictaction = tf.argmax(self.out) self.nextQ = tf.placeholder(shape=[1, 2], dtype=tf.float32) loss = tf.reduce_sum(tf.square(self.out - self.nextQ)) trainer = tf.train.GradientDescentOptimizer(learning_rate=0.1) self.updatemodel = trainer.minimize(loss) init = tf.global_variables_initializer() self.sess = tf.Session() self.sess.run(init) def get_N(self): return self.N def __keyWithMinVal(self, d): v = list(d.values()) k = list(d.keys()) return k[v.index(min(v))] def __discountedReward(self, reward): discounted_reward = np.zeros(len(reward)) rsum = 0 for t in reversed(range(0, len(reward))): rsum = self.discount * rsum + reward[t] discounted_reward[t] = rsum return discounted_reward def __getRegret(self): cache = set(self.cachebuff.getleft()) requestSequence = list(self.hist) ## Compute distance dist = {} for j, p in enumerate(requestSequence): if p not in dist: dist[p] = dequecustom() dist[p].append(j) discountedregret = 0 i = 0 batchid = 0 optsum = 0 hitsum = 0 for hits, sz in zip(self.pageHitBuff, self.batchsizeBuff): opthits = 0 batchid += 1 for _ in range(0, sz): p = requestSequence[i] i += 1 if p in cache: opthits += 1 else: if len(cache) >= self.N: rem = 'xxxxxxxxxxxxx' for c in cache: if c not in dist or len(dist[c]) == 0: rem = c break if rem not in dist or dist[c].getleft( ) > dist[rem].getleft(): rem = c ## Evict from cache cache = cache - {rem} ## Add page to cache cache = cache | {p} ## Pop from dist dist[p].popleft() regret = opthits - hits discountedregret = discountedregret + regret * (0.9)**(batchid - 1) optsum += opthits hitsum += hits break return discountedregret def getState(self): x = np.zeros(self.N, np.float32) for i, page in enumerate(self.disk): x[i] = 1.0 * self.freq[page] if np.sum(x) > 0.00001: x = x / np.sum(x) return x ######################################################################################################################################## ####REQUEST############################################################################################################################# ######################################################################################################################################## def request(self, page): page_fault = False ############################ ## Save data for training ## ############################ if len(self.uniquePages) == 0: ## Compute regret for the first batch if len(self.Xbuff) >= self.numbatch: r = self.__getRegret() cache = self.cachebuff.popleft() s1 = np.array(self.Xbuff.popleft()) s2 = np.array(self.Xbuff.getleft()) act = self.Ybuff.popleft() hits = self.pageHitBuff.popleft() sz = self.batchsizeBuff.popleft() for _ in range(0, sz): temp = self.hist.popleft() ############################################################################################################################# ## Train here ############################################################################################################### ############################################################################################################################# allq = self.sess.run(self.out, feed_dict={self.input: s1}) nextq = self.sess.run(self.out, feed_dict={self.input: s2}) Qmax = np.max(nextq) targetQ = allq targetQ[0, act[0]] = r + self.discountrate * Qmax _ = self.sess.run(self.updatemodel, feed_dict={ self.input: s1, self.nextQ: targetQ }) #self.error = self.error * self.reduceErrorRate ##################### ## Choose randomly ## ##################### state = np.array([self.getState()]) #print(state) self.action = self.sess.run(self.predictaction, feed_dict={self.input: state}) if np.random.rand() < self.error: self.action[0] = 0 if np.random.rand() < 0.5 else 1 self.cachebuff.append(self.disk.getData()) self.Xbuff.append(state) self.Ybuff.append(self.action) ######################### ## Process page reques ## ######################### if self.disk.inDisk(page): self.disk.moveBack(page) self.freq[page] += 1 self.currentPageHits += 1 else: if self.disk.size() == self.N: if self.action[0] == 0: ## Remove LRU page lru = self.disk.getIthPage(0) self.disk.delete(lru) del self.freq[lru] elif self.action[0] == 1: ## Remove LFU page lfu = self.__keyWithMinVal(self.freq) self.disk.delete(lfu) del self.freq[lfu] # Add page to the MRU position self.disk.add(page) self.freq[page] = 1 page_fault = True #self.uniquePages = self.uniquePages | {page} self.uniquePages.update({page: 1}) ## Store page hits for current batch if len(self.uniquePages) == self.N: self.pageHitBuff.append(self.currentPageHits) self.batchsizeBuff.append(sum(self.uniquePages.values())) ## Reset variables self.uniquePages.clear() self.currentPageHits = 0 self.hist.append(page) return page_fault def get_data(self): # data = [] # for i,p,m in enumerate(self.T): # data.append((p,m,i,0)) # return data return [self.disk.get_data()] def get_list_labels(self): return ['L']
class LaCReME_context1(page_replacement_algorithm): def __init__(self, N): self.N = N self.CacheRecency = Disk(N) self.CacheFrequecy = priorityqueue(N) self.Hist1 = Disk(N) self.Hist2 = priorityqueue(N) ## Config variables self.decayRate = 1 self.epsilon = 0.90 self.lamb = 0.05 self.learning_phase = N/2 self.error_discount_rate = (0.005)**(1.0/N) # self.error_discount_rate = 1 ## self.learning = True self.policy = 0 self.evictionTime = {} self.policyUsed = {} self.weightsUsed = {} self.freq = {} ## TODO add decay_time and decay_factor self.decay_time = N self.decay_factor = 1 ## Accounting variables self.time = 0 self.W = np.zeros((10,2)) self.W[:,:] = 0.5 self.X = np.array([],dtype=np.int32) self.Y1 = np.array([]) self.Y2 = 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)) l1, = plt.plot(self.X,self.Y1, 'b-', label='W_lru') l2, = plt.plot(self.X,self.Y2, 'r-', label='W_lfu') l3, = plt.plot(self.X, self.NewPages, 'g-', label='New Pages') return [l1,l2,l3] def __keyWithMinVal(self,d): v=list(d.values()) k=list(d.keys()) return k[v.index(min(v))] def getMinValueFromCache(self, values): minpage,first = -1, True for q in self.Cache : if first or values[q] < values[minpage] : minpage,first=q,False return minpage ########################################## ## Add a page to cache using policy 'poly' ########################################## def addToCache(self, page,pagefreq=0): self.CacheRecency.add(page) self.CacheFrequecy.add(page) self.CacheRecency.increaseCount(page, amount=pagefreq) self.CacheFrequecy.increase(page, amount=pagefreq) ###################### ## Get LFU or LFU page ###################### def selectEvictPage(self, policy): r = self.CacheRecency.getIthPage(0) f = self.CacheFrequecy.peaktop() 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): self.CacheRecency.delete(pg) self.CacheFrequecy.delete(pg) ############################################################## ## There was a page hit to 'page'. Update the data structures ############################################################## def pageHitUpdate(self, page): if page in self.CacheRecency and page in self.CacheFrequecy: self.CacheRecency.moveBack(page) self.CacheRecency.increaseCount(page) self.CacheFrequecy.increase(page) ######################### ## Get the Q distribution ######################### def getQ(self,wid): return (1-self.lamb) * self.W[wid] + self.lamb*np.ones(2)/2 ############################################ ## Choose a page based on the q distribution ############################################ def chooseRandom(self, wid): q = self.getQ(wid) r = np.random.rand() for i,p in enumerate(q): if r < p: return i return len(q)-1 def updateWeight(self, cost): self.W = self.W * (1-self.epsilon * cost) self.W = self.W / np.sum(self.W) ######################################################################################################################################## ####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 partitions = 5 newpagesratio = 1.0*self.sum / self.N weigthid = int(newpagesratio*partitions) if newpagesratio < 1.0 else partitions-1 ##################### ## Visualization data ##################### prob = self.getQ(weigthid) self.X = np.append(self.X, self.time) self.Y1 = np.append(self.Y1, prob[0]) self.Y2 = np.append(self.Y2, prob[1]) notInHistory = 0 ########################## ## Process page request ########################## if page in self.CacheFrequecy: page_fault = False self.pageHitUpdate(page) else : ##################################################### ## Learning step: If there is a page fault in history ##################################################### pageevict, histpage_freq = None,1 policyUsed = -1 if page in self.Hist1: pageevict = page histpage_freq = self.Hist1.getCount(page) self.Hist1.delete(page) policyUsed = 0 elif page in self.Hist2: pageevict = page histpage_freq = self.Hist2.getFreq(page) ## Get the page frequency in history self.Hist2.delete(page) policyUsed = 1 else: notInHistory = 1 if pageevict is not None : q = self.weightsUsed[pageevict] err = self.error_discount_rate ** (self.time - self.evictionTime[pageevict]) reward = np.array([0,0], dtype=np.float32) if policyUsed == 0 : # LRU reward[1] = err if policyUsed == 1: reward[0] = err reward_hat = reward / q ################# ## Update Weights ################# if self.policyUsed[pageevict] != -1 : self.W[weigthid] = self.W[weigthid] * np.exp(self.lamb * reward_hat / 2) self.W[weigthid] = self.W[weigthid] / np.sum(self.W[weigthid]) #################### ## Remove from Cache #################### if self.CacheRecency.size() == self.N: ################ ## Choose Policy ################ if not self.learning : act = np.argmax(self.W[weigthid]) else : act = self.chooseRandom(weigthid) cacheevict,poly = self.selectEvictPage(act) pagefreq = self.CacheFrequecy.getCount(cacheevict) self.policyUsed[cacheevict] = poly self.weightsUsed[cacheevict] = self.getQ(weigthid) self.evictionTime[cacheevict] = self.time if not self.learning : self.policyUsed[cacheevict] = -1 ################### ## Evict to history ################### histevict = None if (poly == 0) or (poly==-1 and np.random.rand() <0.5): if self.Hist1.size() == self.N : histevict = self.Hist1.getIthPage(0) self.Hist1.delete(histevict) self.Hist1.add(cacheevict) self.Hist1.setCount(cacheevict, pagefreq) else: if self.Hist2.size() == self.N : histevict = self.Hist2.popmin() self.Hist2.add(cacheevict) self.Hist2.increase(cacheevict, pagefreq-1) if histevict is not None : del self.evictionTime[histevict] del self.policyUsed[histevict] del self.weightsUsed[histevict] self.evictPage(cacheevict) self.addToCache(page, pagefreq=histpage_freq) page_fault = True self.q.put(notInHistory) self.sum += notInHistory if self.q.qsize() > self.N: self.sum -= self.q.get() self.NewPages.append(newpagesratio) return page_fault def get_list_labels(self) : return ['L']
class LaCReME_simple(page_replacement_algorithm): def __init__(self, N): self.N = N self.CacheRecency = Disk(N) self.CacheFrequecy = priorityqueue(N) self.Hist1 = Disk(N) self.Hist2 = priorityqueue(N) ## Config variables self.decayRate = 1 self.epsilon = 0.05 self.learning_phase = N / 2 ## self.learning = True self.policy = 0 self.evictionTime = {} self.policyUsed = {} self.weightsUsed = {} self.freq = {} ## TODO add decay_time and decay_factor self.decay_time = N self.decay_factor = 1 ## Accounting variables self.time = 0 self.W = np.array([.5, .5], dtype=np.float32) ## For visualization self.X = np.array([], dtype=np.int32) self.Y1 = np.array([]) self.Y2 = np.array([]) def get_N(self): return self.N def visualize(self, plt): ax = plt.subplot(2, 1, 1) ax.set_xlim(np.min(self.X), np.max(self.X)) l1, = plt.plot(self.X, self.Y1, 'b-', label='W_lru') l2, = plt.plot(self.X, self.Y2, 'r-', label='W_lfu') return [l1, l2] ########################################## ## Add a page to cache using policy 'poly' ########################################## def addToCache(self, page, pagefreq=0): self.CacheRecency.add(page) self.CacheFrequecy.add(page) self.CacheRecency.increaseCount(page, amount=pagefreq) self.CacheFrequecy.increase(page, amount=pagefreq) ###################### ## Get LFU or LFU page ###################### def selectEvictPage(self, policy): r = self.CacheRecency.getIthPage(0) f = self.CacheFrequecy.peaktop() 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): self.CacheRecency.delete(pg) self.CacheFrequecy.delete(pg) ############################################################## ## There was a page hit to 'page'. Update the data structures ############################################################## def pageHitUpdate(self, page): if page in self.CacheRecency and page in self.CacheFrequecy: self.CacheRecency.moveBack(page) self.CacheRecency.increaseCount(page) self.CacheFrequecy.increase(page) ############################################ ## Choose a page based on the q distribution ############################################ def chooseRandom(self): q = self.getQ() r = np.random.rand() for i, p in enumerate(q): if r < p: return i return len(q) - 1 ################## ####REQUEST####### ################## def request(self, page): page_fault = False self.time = self.time + 1 ##################### ## Visualization data ##################### self.X = np.append(self.X, self.time) self.Y1 = np.append(self.Y1, self.W[0]) self.Y2 = np.append(self.Y2, self.W[1]) if self.time % self.N == 0: self.CacheFrequecy.decay(self.decay_factor) self.Hist2.decay(self.decay_factor) ########################## ## Process page request ########################## if page in self.CacheFrequecy: page_fault = False self.pageHitUpdate(page) else: ##################################################### ## Learning step: If there is a page fault in history ##################################################### histpage_freq = 1 if page in self.Hist1: histpage_freq = self.Hist1.getCount(page) self.Hist1.delete(page) self.W[0] = self.W[0] * ( 1 - self.epsilon) if self.policyUsed[page] != -1 else self.W[0] elif page in self.Hist2: histpage_freq = self.Hist2.getCount( page) ## Get the page frequency in history self.Hist2.delete(page) self.W[1] = self.W[1] * ( 1 - self.epsilon) if self.policyUsed[page] != -1 else self.W[1] #################### ## Normalize weights #################### self.W = self.W / np.sum(self.W) #################### ## Remove from Cache #################### if self.CacheRecency.size() == self.N: ################################################################ ## Choose Policy by picking the one with the highest weight ################################################################ act = np.argmax(self.W) cacheevict, poly = self.selectEvictPage(act) pagefreq = self.CacheFrequecy.getCount(cacheevict) self.policyUsed[cacheevict] = poly self.weightsUsed[cacheevict] = self.W self.evictionTime[cacheevict] = self.time ################### ## Evict to history ################### histevict = None if (poly == 0) or (poly == -1 and np.random.rand() < 0.5): if self.Hist1.size() == self.N: histevict = self.Hist1.getIthPage(0) self.Hist1.delete(histevict) self.Hist1.add(cacheevict) self.Hist1.setCount(cacheevict, pagefreq) else: if self.Hist2.size() == self.N: histevict = self.Hist2.popmin() self.Hist2.add(cacheevict) self.Hist2.increase(cacheevict, pagefreq - 1) ################################## ## Remove histevict from the system ################################### if histevict is not None: del self.evictionTime[histevict] del self.policyUsed[histevict] del self.weightsUsed[histevict] ############################################## ## Remove cacheevict from both data structures ############################################## self.evictPage(cacheevict) ################################### ## Add page to both data structures ################################### self.addToCache(page, pagefreq=histpage_freq) page_fault = True return page_fault def get_list_labels(self): return ['L']
class ExpertLearning(page_replacement_algorithm): def __init__(self, N): self.T = [] self.N = N self.disk = Disk(N) self.freq = {} ## Training variables self.X, self.Y = [], [] self.reward = [] self.regret = [] ## Config variables self.batchsize = N self.numbatch = 5 ## Aux variables self.hist = queue.deque() self.Xbuff = queue.deque() self.Ybuff = queue.deque() self.pageHitBuff = deque() self.current = 0 self.action = 1 self.currentPageHits = 0 #self.discount = 0.9 #self.sampleCount = 0 #self.trainingSampleSize = 5 * N def get_N(self): return self.N def __keyWithMinVal(self, d): v = list(d.values()) k = list(d.keys()) return k[v.index(min(v))] def __discountedReward(self, reward): discounted_reward = np.zeros(len(reward)) rsum = 0 for t in reversed(range(0, len(reward))): rsum = self.discount * rsum + reward[t] discounted_reward[t] = rsum return discounted_reward def __getRegret(self): return 0 def getState(self): x = np.zeros(self.N, np.float32) for i, page in enumerate(self.disk): x[i] = 1.0 * self.freq[page] if np.sum(x) > 0.00001: x = x / np.sum(x) return x def request(self, page): page_fault = False ############################ ## Save data for training ## ############################ if self.current == 0: ## Compute regret for the first batch if len(self.hist) == self.numbatch * self.batchsize: reg = self.__getRegret() ## Regret of first n pages x = self.Xbuff.popleft() y = self.Ybuff.popleft() h = self.pageHitBuff.popleft() ## Remove from hist and buffers for _ in range(0, self.N): self.hist.get() ## Choose randomly self.action = 1 if np.random.rand() < 0.5 else 2 self.Xbuff.append(self.getState()) self.Ybuff.append(self.action) ######################### ## Process page reques ## ######################### if self.disk.inDisk(page): self.disk.moveBack(page) self.freq[page] += 1 else: if self.disk.size() == self.N: if self.action == 1: ## Remove LRU page lru = self.disk.getIthPage(0) self.disk.delete(lru) del self.freq[lru] elif self.action == 2: ## Remove LFU page lfu = self.__keyWithMinVal(self.freq) self.disk.delete(lfu) del self.freq[lfu] # Add page to the MRU position self.disk.add(page) self.freq[page] = 1 page_fault = True ## Increate page hits counter self.currentPageHits += 1 * (not page_fault) ## Save page hits for current batch if self.current + 1 == self.batchsize: self.pageHitBuff.append(self.currentPageHits) ## Save page in history self.hist.put(page) ## Increase batch size counter self.current = (self.current + 1) % self.batchsize return page_fault def get_data(self): # data = [] # for i,p,m in enumerate(self.T): # data.append((p,m,i,0)) # return data return [self.disk.get_data()] def get_list_labels(self): return ['L']