We investigate a dynamic, adaptive resource allocation scheme with the aim of accelerating the convergence of multi-start population-based search heuristics (PSHs) running on multiple parallel processors. Given that each initialization of a PSH performs differently over time, we develop an exponential learning scheme which allocates computational resources (processors) to each variant in an online manner, based on the performance level attained by each initialization. For the well-known example of (mu+lambda)-evolution strategies, we show that the time required to reach the target quality level of a given optimization problem is significantly reduced and that the utilization of the parallel system is likewise optimized. Our learning approach is easily implementable with currently available batch management systems and provides notable performance improvements without modifying the employed PSH, so it is very well-suited to improve the performance of PSHs in large-scale parallel computing environments.

Parallel task graphs (PTGs) arise when parallel programs are combined to larger applications, e.g., scientific workflows. Scheduling these PTGs onto clusters is a challenging problem due to the additional degree of parallelism stemming from moldable tasks. Most algorithms are based on the assumption that the execution time of a parallel task is monotonically decreasing as the number of processors increases. But this assumption does not hold in practice since parallel programs often perform better if the number of processors is a multiple of internally used block sizes. In this article, we introduce the Evolutionary Moldable Task Scheduling (EMTS) algorithm for scheduling static PTGs onto homogeneous clusters. We apply an evolutionary approach to determine the processor allocation of each task. The evolutionary strategy ensures that EMTS can be used with any underlying model for predicting the execution time of moldable tasks. With the purpose of finding solutions quickly, EMTS considers results of other heuristics (e.g., HCPA, MCPA) as starting solutions. The experimental results show that EMTS significantly reduces the make span of PTGs compared to other heuristics for both non-monotonically and monotonically decreasing models.

This paper proposes a comprehensive modeling architecture for workloads on parallel computers using Markov chains in combination with state dependent empirical distribution functions. This hybrid approach is based on the requirements of scheduling algorithms: the model considers the four essential job attributes submission time, number of required processors, estimated processing time, and actual processing time. So far, no model exists that considers all those attributed at the same time. To assess the goodness-of-fit of a workload model the similarity between sequences of real jobs and jobs generated from the model needs to be captured. We propose to reduce the complexity of this task and to evaluate the model by comparing the results of a widely-used scheduling algorithm instead. This approach is demonstrated with commonly used scheduling objectives. To verify this evaluation technique, standard criteria for assessing the goodness-of-fit for workload models are additionally applied.

In this paper, we address the problem of finding well-performing workload exchange policies for decentralized Computational Grids using an Evolutionary Fuzzy System. To this end, we establish a noninvasive collaboration model on the Grid layer which requires minimal information about the participating High Performance and High Throughput Computing (HPC/HTC) centers and which leaves the local resource managers completely untouched. In this environment of fully autonomous sites, independent users are assumed to submit their jobs to the Grid middleware layer of their local site, which in turn decides on the delegation and execution either on the local system or on remote sites in a situation-dependent, adaptive way. We find for different scenarios that the exchange policies show good performance characteristics not only with respect to traditional metrics such as average weighted response time and utilization, but also in terms of robustness and stability in changing environments.

In our work, we utilize a competitive Co-evolutionary Algorithm in order to optimize the parameter set of a Fuzzy System for job exchange in Computational Grids. In this domain, the providers of High Performance Computing (HPC) centers strive for minimizing the response time for their own customers by trying to distribute workload to other sites in the Grid environment. The Fuzzy System is used for steering each site's decisions whether to distribute or accept workload in a beneficial, yet egoistic direction. This scenario is particularly suited for the application of a competitive CA: Grid sites' Fuzzy Systems are modeled as species, which evolve in different populations. While each species tries to minimize the response time for locally submitted jobs, their individuals' fitness is determined within the commonly shared ecosystem. Using real workload traces and Grid setups, we show that the opportunistic cooperation leads to significant improvements for both each Grid site and the overall system.

In our work, we address the problem of workload distribution within a computational grid. In this scenario, users submit jobs to local high performance computing (HPC) systems which are, in turn, interconnected such that the exchange of jobs to other sites becomes possible. Providers are able to avoid local execution of jobs by offering them to other HPC sites. In our implementation, this distribution decision is made by a fuzzy system controller whose parameters can be adjusted to establish different exchange behaviors. In such a system, it is essential that HPC sites can only benefit if the workload is equitably (not necessarily equally) portioned among all participants. However, each site egoistically strives only for the minimization of its own jobs' response times regularly at the expense of other sites. This scenario is particularly suited for the application of a competitive coevolutionary algorithm: the fuzzy systems of the participating HPC sites are modeled as species that evolve in different populations while having to compete within the commonly shared ecosystem. Using real workload traces and grid setups, we show that opportunistic cooperation leads to significant improvements for each HPC site as well as for the overall system.

This paper examines the job exchange between parallel compute sites in a decentralized grid scenario. Here, the local scheduling system remains untouched and continues normal operation. In order to establish the collaboration and interaction between sites in a grid context, a middleware layer that is responsible for the migration of jobs is supplemented. Independent users are assumed to submit their jobs to their site-local middleware layer, which in turn can request jobs for execution from alien sites. The simulation results are obtained using real workload traces and compared to the performance of the EASY Backfilling algorithm in an equal single-site scenario. It is shown that collaboration between site is beneficial for all high utilized participants as it is possible to achieve shorter response times for jobs compared to the best single-site scheduling results.

In this paper, we introduce a methodology for the approximation of optimal solutions for a resource allocation problem in the domain of Grid scheduling on High Performance Computing systems. In detail, we review a real-world scenario with decentralized, equitable, and autonomously acting suppliers of compute power who wish to collaborate in the provision of their resources. We exemplarily apply NSGA-II in order to explore the bounds of maximum achievable benefit. To this end, appropriate encoding schemes and variation operators are developed while the performance is evaluated. The simulations are based upon recordings from real-world Massively Parallel Processing systems that span a period of eleven months and comprise approximately 100,000 jobs. By means of the obtained Pareto front we are able to identify bounds for the maximum benefit of Grid computing in a popular scenario. For the first time, this enables Grid scheduling researchers to rank their developed real-world strategies.

This paper presents a comparison of three different design concepts for genetic fuzzy systems. We apply a symbiotic evolution that uses the Michigan approach and two approaches that are based on the Pittsburgh approach: a complete optimization of the problem and a cooperative coevolutionary algorithm. The three different genetic fuzzy systems are applied to a real-world online problem, the generation of scheduling strategies for massively parallel processing systems. The genetic fuzzy systems must classify different scheduling states and decide about a corresponding scheduling strategy within each scheduling state. The main challenge arise in the delayed reward given by a critic. Therefore, it is impossible to directly evaluate the assignment of scheduling strategies to scheduling states. In our paper, the three design concepts are evaluated with real workload traces considering result quality, computational effort, convergence behavior, and robustness.

In this paper, we utilize a predator-prey model in order to identify characteristics of single-objective variation operators in the multi-objective problem domain. In detail, we analyze exemplarily Gaussian mutation and simplex recombination to find explanations for the observed behaviorswithin this model. Then, both operators are combinedto a new complex one for the multi-objective case in order to aggregate the identified properties. Finally, we show that (a) characteristic properties can still be observed in the combination and (b) the collaboration of those operators is beneficial for solving an exemplary multi-objective problem regarding convergence and diversity.