Performance results on SAGrid for libsvm training and characterisation studies

The Support Vector Machine library has been used together with sets of data collected during the Lwazi programme at the Meraka Institute, in order to estimate the performance of the South African National Grid (SAGrid) sites in dealing with typical machine-learning tasks. Processing of standard data sets of differing sizes was done on various sites on the infrastructure, and the time needed to conclude tasks such as model training and characterisation measured. This dataset is the timing of these tasks, for varying levels of task parallelism, in order to determine the real-world performance of this application, as well as validate expectations from running similar workflows on centralised resources. This was collected in the context of one of the authors' (D. Risinamhodzi) M.Sc. thesis. Furthermore, this dataset implicitly demonstrates the functionality and performance of the CODE-RADE platform.

本研究采用支持向量机（Support Vector Machine）库，结合梅拉卡研究所（Meraka Institute）Lwazi项目采集的数据集，以评估南非国家网格（South African National Grid, SAGrid）各节点在处理典型机器学习任务时的性能表现。研究团队在该网格基础设施的多个节点上运行了不同规模的标准数据集处理任务，并对模型训练、特征表征等任务的完成所需时长进行了测量。本数据集记录了不同任务并行度下上述所有任务的耗时，用于评估该应用的实际运行性能，同时验证在集中式资源上运行同类工作流的预期效果。该数据集的采集工作源于作者之一D. Risinamhodzi的硕士学位论文研究。此外，本数据集还间接展现了CODE-RADE平台的功能与性能表现。