Computing Science: Predicting the Execution Time of Workflow Blocks Based on Their Input Features.

摘要

The ability to accurately estimate the execution time of computationally expensive escience algorithms enables better scheduling of workflows that incorporate those algorithms as their building blocks, and may give users an insight into the expected cost of workflow execution on cloud resources. When a large history of past runs can be observed, crude estimates such as the average execution time can easily be provided. We make the hypothesis that, for some algorithms, better estimates can be obtained by using the histories to learn regression models that predict execution time based on selected features of their inputs. We refer to this property as input predictability of algorithms. We are motivated by e-science workflows that involve repetitive training of multiple learning models. Thus, we verify our hypothesis on the specific case of the C4.5 decision tree builder, a well-known learning method whose training execution time is indeed sensitive to the specific input dataset, but in nonobvious ways. We use the case study to demonstrate a method for assessing input predictability. While this yields promising results, we also find that its more general applicability involves a trade off between the black-box nature of the algorithms under analysis, and the need for expert insight into relevant features of their inputs.