Solution Biasing for Optimized Cloud Workload Placement Asser N. Tantawi 2016 IEEE International Conference on Autonomic Computing Summarized by Uma Tadakamalla 1
I. Introduction Outline II. III. Problem Statement Solution IV. Functional Definitions V. Experimental Results 2
I. Introduction Problem domain: cloud computing Problem: workload placement in the cloud 3
II. Problem Statement Mapping of Logical Entities (LE) onto Physical Entities (PE) in such a way as to Optimize an Objective Function subject to Constraints 4
II. Problem Statement, cont. Logical Entities (LE): Virtual Machines (VM), Data Volumes, Communication links, and Containers Physical Entities (PE): Physical Machines (PM), Storage Devices, and Communication networks Objective Function: User and Provider Objectives Constraints: Properties and/or Resource Demands, Communications needs, Collocation(or anticollocation) of member LEs of the group 5
II. Problem Statement, cont. N physical entities, M logical entities X is a variable mapping logical to physical entities S the set of possible solutions satisfying constraints Objective function: f(x) A is a particular mapping (solution) 6
III. Solution Optimization problem is formulated as a General Search Problem based on Random Search Methods Uses biasing and sampling Does not depend on properties such as convexity or heuristics 7
III. Solution, cont. Generalized Random Search Method 8
III. Solution, cont. Biased Sampling Algorithm (BSA) a random search algorithm uses sampling and biasing makes the following choices: Neighborhood H(k) is characterized by the marginal probability distributions Candidate solution X (k) is generated using a modified Gibbs sampling Next point X(k + 1) is the generated point X (k) Estimate of optimal solution X (k) is generated using importance sampling technique. 9
III. Solution, cont. Biased Sampling Algorithm (BSA) 10
Stopping criteria could be: number of iterations or III. Solution, cont. reaching diminishing return After the stopping criterion is met, the sample with the minimum f() is the solution. Time complexity is O(NM 2 ). 11
IV. Functional Definitions Biasing helps in finding good solutions in less time Biasing functions push the solution towards one that satisfies the constraints and optimizes the objective function. Biasing functions are designed to have parameters, also known as bias factors. Setting the values for bias factors appropriately is critical to the quality of the results. 12
IV. Functional Definitions, cont. Usage Biasing: non-increasing function of resource utilization Location Biasing: non-increasing function in the amount of deviation from the location preferences set by the user 13
IV. Functional Definitions, cont. Objective. Function: weighted sum of system (provider) objective and pattern (user) objective System objective: standard deviation of the utilization of resources The objective is to balance the load across the cloud Pattern objective: deviation from the desired location constraints specified in the pattern request 14
V. Experimental Results Cloud System Configuration (simulated) 15
V. Experimental Results, cont. Three experiments were conducted with several pattern configurations to: demonstrate the efficiency of the BSA algorithm investigate the sensitivity of the bias factors. 16
V. Experimental Results (Exp #1) 17
V. Experimental Results (Exp #1) Conducted several experiments with β usg = [0, 8] The rate of improvement decreases from an initial 0.0192, going from 0 to 0.5, to an insignificant 0.0017, going from 6 to 8. Used a default value of β usg = 2 The algorithm execution time is < 100 msec 18
V. Experimental Results (Exp #1) 19
V. Experimental Results (Exp #1) Experiment #1: Optimality Test Optimal Policy: a VM is placed in the PM with the smallest CPU utilization The results of the BSA algorithm are within the 20% of optimum solution obtained using the best-fit policy. 20
V. Experimental Results (Exp #2) 21
V. Experimental Results (Exp #2) Conducted several experiments with β loc = [2, 6] As β loc increases, the pattern value decreases, and at the same time, system value increases At β loc = 4, the total value is a minimum (one of the goals) The algorithm execution time ~ 200 msec 22
V. Experimental Results (Exp #2) 23
V. Experimental Results (Exp #3) 24
V. Experimental Results (Exp #3) 25
V. Experimental Results (Exp #3) Conducted several experiments with β loc = [0, 8] As β usg increases, the system value decreases and the pattern value increases, resulting in a fairly constant total value around 0.17 The algorithm execution time ~ 25 msec. 26
Conclusions & Critique 1. The BSA algorithm appears to be fast -- finds solutions in 25 to 200 msec for the experiments conducted. 2. Author claimed: In a case where an optimal solution may be obtained, we show that our method is within 20% of optimality The above happens to be for a very simple case, i.e., Exp. #1 How can one be sure that is case for a more general scenarios with user objective function and/or constraints? 3. The BSA algorithm has not been compared (either time complexity or accuracy) with state of the art algorithms! 27