Grid computing exploits the availability of many heterogeneous processing elements connected by means of the Internet to achieve high-performance computing for various large-scale application environments. One of several important issues in this scenario is the design of a resource management system that at any given time distributes the load among the available resources--each one with potentially heterogeneous characteristics--to maximize performance. The configuration of the system changes dynamically based on the nodes that become available and the ones that go offline. Eventually, a failure of one or more of the grid processing elements may occur, and it may be either a transient problem or a permanent one. Therefore, a robust load balancing strategy should take into account such a possible scenario and be able to address the detection of failures and the need to distribute the load to avoid the failed resources.
In recent decades, a lot of work has addressed load balancing. The authors provide a critical overview of the literature in order to identify the elements to be exploited and the limitations to be overcome; they also introduce some peculiarities. The grid computing environment shares a lot of commonalities with past application scenarios and architectures for which load balancing strategies have been defined (for example, distributed architectures).
The model of all the relevant elements (for example, system, application, and fault) offers a complete view of the working scenario and sets the basis for the presentation of the proposed load balancing policy, which exhibits both static and dynamic characteristics to achieve the fault-tolerant resource-usage optimization goal while fulfilling the job execution deadline constraints. The static policy comes from the literature and refers to existing solutions, whereas the dynamic one constitutes an innovative contribution. The authors introduce two key aspects: information (to keep track of the status of the processing elements) and load distribution policies. The fault management approach is based on replication.
The outcome is a policy--namely, AlgHybrid_LB--that promises to outperform existing classical solutions such as minimum completion time (MCT) and Min-min. To support such a claim, the authors provide a broad set of experimental results performed in different working scenarios (for example, in terms of the number of available resources), showing significant improvements with respect to the analyzed parameters. The results are thoroughly described and reported at a high abstraction level, so the impact of the various aspects cannot be fully perceived; they offer only an overview of the final high-level performance.
The paper is well organized and easy to follow, even for those who are not grid computing experts; in fact, the proposed approach seems to be easily adjustable to similar scenarios. Indeed, the paper offers all the information necessary so that readers can re-implement the discussed solution as a starting point for future developments or for comparing alternatives, which is important when considering the value of a proposal. Finally, the “Related Work” section is an interesting source for the literature.
