Multi-Objective Problem Solving With Offspring on Enterprise Clouds Christian Vecchiola, Michael Kirley, and Rajkumar Buyya Department of Computer Science & Software Engineering The University of Melbourne, 3053, Carlton, Victoria, Australia {csve, mkirley, raj}@csse.unimelb.edu.au Abstract problems with a large number of objectives. Initially, we restricted the implementation to a sequential model. In this paper, we present a distributed implementa- In this paper, we introduce its distributed implemen- tion of a network based multi-objective evolutionary tation by using Offspring, which is a framework we algorithm called EMO by using Offspring. Network developed for distributing the execution of evolutionary based evolutionary algorithms have proven to be effec- algorithms. Enterprise Clouds [5] provide the required tive for multi-objective problem solving. They feature a computational power to solve large optimization prob- network of connections between individuals that drives lems in a reasonable period. Offspring provides facili- the evolution of the algorithm. Unfortunately, they re- ties for distributing the large computation load gener- quire large populations to be effective and a distrib- ated by MOEAs, by simply asking the user to define the uted implementation can leverage the computation time. strategy to use for coordinating the distributed execu- Most of the existing frameworks are limited to provid- tion. The primary aim of this system is to provide a ing solutions that are basic or specific to a given algo- friendly user environment for researchers in combinato- rithm. Our Offspring framework is a plug-in based rial optimization who do not want to be concerned software environment that allows rapid deployment about building interconnection software layers and and execution of evolutionary algorithms on distrib- learning underlying middleware APIs. Specifically, we uted computing environments such as Enterprise provide a visual user interface that manages the execu- Clouds. Its features and benefits are presented by de- tion of population based optimization algorithms, a set scribing the distributed implementation of EMO. of APIs allowing researchers to write a plug-in for this environment quickly. The distributed version of our 1. Introduction serial implementation of MOEA has been developed as a plug-in for this system by simply defining a strategy which: (i) coordinates the different serial executions Many problems in science, engineering, and economics distributed among the nodes; and (ii) applies smart mi- require solutions consisting of several incommensur- grations at the end of each of the iterations of the algo- able and possibly conflicting objectives, constraints, rithm. Even though the infrastructure provided by Off- and/or problem parameters. Multi-objective evolution- spring is general enough to deploy a distributed imple- ary algorithms (MOEAs) are now a well-established mentation of any population based algorithms, the real population based metaheuristic used to find a set of advantage is obtained when a distributed implementa- Pareto-optimal solutions for such problems [1]. How- tion is composed by coordinating the runs of the serial ever, one of the major difficulties when applying implementation of the same algorithm. In that case, the MOEAs to real-world problems is the computational distributed implementation with Offspring is obtained cost associated with the large number of function with a minimal coding effort, because it is only neces- evaluations necessary to obtain a range of acceptable sary to code the coordination strategy. These conditions solutions. In the MOEA domain, there have only been a apply to the population based metaheuristics making relatively small number of parallel models described as use of topology information to improve the quality of compared with the single objective domain (see Veld- solutions. huizen et al. [2] for a review). Recently, we introduced The rest of the paper is organized as follows. In Sec- a novel complex network-based MOEA [3, 4], called tion 2, we describe the related work in virtualization EMO, to address the many inherent challenges when technologies and distributed metaheuristics. Section 3 attempting to find a range of solutions, particularly for provides a very brief introduction to population based metaheuristics and introduces the challenges in distrib- uting network based evolutionary algorithms. Section 4 each of the nodes. These tasks cannot be performed and 5 describe the architecture of the distributed im- with Nimrod/O that simply provides a technique for plementation of EMO by using Offspring. Some pre- partitioning the problem space and distribute the com- liminary results are reported in Section 6. Conclusions putation. For these reasons, Offspring is more similar to and plans for future work follow in Section 7. DREAM since it provides a distribution engine making the development of distributed evolutionary algorithms 2. Background straightforward. The approach used by DREAM to dis- tribute the computation is based on mobile multi-agent The idea of providing support for distributed execution systems, while Offspring relies on the Enterprise of nature inspired population based metaheuristics has Clouds. Compared to ParadisEO-MOEO Offspring been investigated with interest in the last two decades provides a smaller set of features, especially for what [6]. In particular, this topic has been thoroughly inves- concerns the statistical analysis of the solutions. The tigated for genetic algorithms and different parallel API provided by ParadisEO-MOEO allows developers execution models have been devised [7, 8]. There exist to virtually control any aspect of the implementation of a wide range of grid middleware technologies, such as a distributed metaheuristic. This great flexibility makes Alchemi [9], Condor-G [10], the Globus Toolkit [11], the development of a new metaheuristic not straight- and grid resource brokering technologies, such as Nim- forward, but a good understanding of the APIs is re- rod/G [12] and Grid Service Broker [13], that have quired. The primary concern of Offspring is to provide simplified the development of distributed problem solv- simple and easy to use abstractions allowing research- ing environments. In this work, we focus on developing ers to compose a distributed metaheuristic by giving the Offspring framework in .NET based Cloud Com- them the maximum freedom on the policies used to puting environments. coordinate the distributed execution. As a result, the For what concerns distributed optimization, different number of APIs to learn and use has been kept minimal. solutions are now available for researchers. Nimrod/O Moreover, another feature that distinguishes Offspring [14] is a tool allowing running distributed optimization from the solutions presented is the use of Enterprise problems by using any Nimrod based system, such as Clouds and Computational Grids. Nimrod/G, as distribution infrastructure. Nimrod/O allows users to take advantage of different optimization 3. Distributed Evolutionary Algorithms algorithms (BGFS, Simplex, Divide and Conquer, and Simulated Annealing). It requires users to specify the Evolutionary algorithms are a class of population based structure of the optimization problem and the variable metaheuristics [6] exploiting the concept of population that needs to be minimized. ParadisEO-MOEO [15] is evolution to find solutions to optimization problems. A an object-oriented framework that provides a full fea- population is a collection of individuals where each tured object model for implementing distributed meta- individual represents or maintains information heuristics, by focusing on code reuse and efficiency. It about a specific solution of the optimization problem. supports MPI, Condor-G, and Globus as distributing The optimal solution is then found by using an iterative middleware technologies. DREAM (Distributed Re- process that evolves the collection of individuals in source Evolutionary Algorithm Machine) [16] provides order to improve the quality of the solution. Genetic a software infrastructure and a technology for the Algorithms (GAs) [17] are the most popular evolution- automatic distribution of evolutionary algorithm proc- ary algorithms. They imitate the process by which na- essing. DREAM is based on a virtual machine that uses ture creates new chromosomes by recombining and a P2P mobile agent system for distributing the compu- mutating existing chromosomes in order to generate the tation. Other minor projects such as TEA, JDEAL, and new population. Figure 1 describes the structure of JMETAL mostly focus on providing a good support for these algorithms. metaheuristic implementation and put less emphasis on When tackling real world problems, such as those the integration with distributed computing technologies. described in Handl et al. [18], the compute intensive Nimrod/O provides a technique for distributing a set step is the evaluation of each individual. A range of of built-in optimization algorithms that is based on pa- structured or parallel genetic algorithms has been pro- rameter sweeping. Offspring provides a more general posed where the population is decentralized in some approach and an extensible platform for creating dis- way (see Cantù-Paz [7] and Alba et al. [8] for an over- tributed evolutionary algorithms. With Offspring, re- view). The models may be loosely classified into one of searchers can either define the structure of the distrib- the