Estimating performance and of cloud-based systems: A model driven, complementary approach.

摘要

Cloud computing has increasingly been adopted for various real-life applications, thanks to a number of benefits it is anticipated to bring, such as cost reduction, improved performance, elasticity, scalability, and the like---the so-called non-functional characteristics. In migrating to or adopting cloud computing, it seems important to be able to estimate or predict such non-functional characteristics and yet challenging. A most reliable for doing an estimation or prediction is likely to involve loading all the particular software system (application), together with all the data that it needs on a target cloud platform, configuring and testing the software system, running it and obtaining metrics of interests. However, this approach can be costly, while involving an extensive amount of time. In this thesis, we propose a model-driven, complementary approach to estimating and predicting the cost and performance of applications to run on a cloud platform. It is model-driven in the sense that (conceptual) models, in particular, ontological concepts are used that are the essence of various estimation techniques. It is complementary in the sense, that several estimations techniques are used in a complementary manner, involving benchmarking, simulation, emulation and a genetic algorithm. Various benchmark results have been produced in the past, although the majority are not for cloud computing-related benchmarks, with the hope that the performance of a new application can be estimated through the use of one of the benchmarks whose characteristics are similar to the new applications.;A key issue with the use of most benchmarking results, however, is whether the comparison between a new software application, which might be migrated to a cloud computing platform, and a benchmark makes sense, or instead like comparing apples and oranges. This issue becomes exacerbated, when we use multiple different techniques. In order to tackle this issue, in our complementary approach, we make comparisons in terms mis-matches and similarities between models---e.g., a benchmark model and a simulation model. To see that our model-driven, complementary approach can help predict the cost and performance of a new application reasonably well, we have run various experimentations, including experimentations on TPC-C, which is an industry standard benchmark specification for Online Transaction Processing domain and widely being used, and experimentations on Yahoo! Cloud Serving Benchmark (YCSB), which supports benchmarking cloud systems and NoSQL database.