Research on Virtual Machine Response Time Prediction Method Based on GA-BP Neural Network

Algorithm 1

Component response time prediction procedure.

Input: The most similar virtual machine’s number of component current requests and

   response time historical sequence is 𝑁_𝑖(𝑐₁),𝑁_𝑖(𝑐₂),…,𝑁_𝑖(𝑐_𝑛)𝑇_𝑖(𝑐₁),𝑇_𝑖(𝑐₂),…,𝑇_𝑖(𝑐_𝑛)

    𝑖=1,2,…,𝑡 before time 𝑡. The number of component current requests is 𝑘.

Output: Component response time 𝑇_𝑖(𝑐₁),𝑇_𝑖(𝑐₂),…,𝑇_𝑖(𝑐_𝑛)𝑖=𝑡+1,𝑡+2,…,𝑡+𝑘 in the future 𝑘 time intervals.

(1) Determine the number of concurrent requests, set the related parameters, establish the BP neural network.

(2) Set up population size 𝐿, chromosome length 𝑀, number of iterations 𝑀𝑎𝑥𝑔𝑒𝑛, crossover probability 𝑃_𝑐 and mutation

  probability 𝑃_𝑚. Initialize 𝑀 sets of weighting value of the threshold. Form population.

(3) Train the BP neural network. Calculate individual adaptiveness according to formula (15). Choose the optimal chromosome.

(4) Use roulette wheel selection method to choose 𝑀 the most adaptive individuals.

(5) Choose several pairs of chromosomes in Step  4 and random number 𝑟,𝑟∈[0,1] for every pair. If 𝑟<𝑃_𝑐,

  chromosome execute crossover operation.

(6) Choose several pairs of chromosomes in Step  5 and random number 𝑟,𝑟∈[0,1] for every pair. If 𝑟<𝑃_𝑚,

  chromosome execute mutation operation.

(7) Choose the individual in Step  6 as network weight threshold to train BP network. Compare the optimal individual

  with the last generation optimal individual.

  If the former is more adaptive, it will replace the last generation optimal individual. Moreover, the last generation optimal

  chromosome replace the worst chromosome.

   𝑔𝑒𝑛=𝑔𝑒𝑛+1, if 𝑔𝑒𝑛=𝑀𝑎𝑥𝑔𝑒𝑛, algorithm will be stoped. Then we get optimal weight threshold. Otherwise, go to Step  4.

(8) Decompose the optimal individual to get BP neural network connective weight threshold.

(9) BP network initial iteration value 𝑖=1.

(10) Put the amount of component concurrent requests and response time sequences

  into model, and calculate component response time according to formula (3).

(11) Formula (12) calculates the first sample component response time prediction error

  quadratic sum and formula (9) adjusts weight.

(12) Choose the second sample component current requests amount and response time. Repeat Steps  10 and 11 until

   finishing 𝑁 samples training.

(13) 𝑖=𝑖+1. if 𝑖=𝑁 or the overall error of the network 𝐸<𝜀. When training process is ended,

   we get future component response time. Otherwise, go to Step  10.

(14) Put the number of 𝑘 components current requests into trained network. Then get the new virtual machine

   component 𝑘 response time.