Online Task Resource Consumption Prediction for Scientific Workflows

Abstract

Estimates of task runtime, disk space usage, and memory consumption, are commonly used by scheduling and resource provisioning algorithms to support efficient and reliable workflow executions. Such algorithms often assume that accurate estimates are available, but such estimates are difficult to generate in practice. In this work, we first profile five real scientific workflows, collecting fine-grained information such as process I/O, runtime, memory usage, and CPU utilization. We then propose a method to automatically characterize workflow task requirements based on these profiles. Our method estimates task runtime, disk space, and peak memory consumption based on the size of the tasks' input data. It looks for correlations between the parameters of a dataset, and if no correlation is found, the dataset is divided into smaller subsets using a clustering technique. Task estimates are generated based on the ratio parameter/input data size if they are correlated, or based on the probability distribution function of the parameter. We then propose an online estimation process based on the MAPE-K loop, where task executions are monitored and estimates are updated as more information becomes available. Experimental results show that our online estimation process results in much more accurate predictions than an offline approach, where all task requirements are estimated prior to workflow execution.

Using simple PID-inspired controllers for online resilient resource management of distributed scientific workflows

“…Typically, mean values yield high values of standard deviation (due to variations inherent to the application itself, or the system including external load), thus estimations may not be accurate. Task characteristics estimation is beyond the scope of this work, and sophisticated methods to provide accurate estimates can be found in [22,23,24,25]. However, this work intends to demonstrate that even using inaccurate estimation methods, our proposed process can cope with the poor estimates and still yield good results.…”

“…Although several works address task requirement estimations based on provenance data [22,23,24,25], accurate estimations are still challenging, and may be specific to a certain type of application. In [26], a prediction algorithm based on machine learning (Naïve Bayes classifier) is proposed to identify faults before they occur, and to apply preventive actions to mitigate the faults.…”

“…Therefore, the decisions, actions, and considered metrics have to remain simple and to yield good results while the application is still executing. Second, it is unpredictable due to the lack of reliable application and platform models [22], and due to the lack of information about the performance of computing and network resources in production environments. Hence, platform and application models also have to remain simple, and adapt to the dynamic behavior of the system.…”

“…The Decision Agent may also be based on an Intelligent System, where decisions do also account, for example, for historical data, system performance metrics (e.g., I/O or network throughput, etc. ), workflow structure and lookahead planning [11,36], and the use of statistical and machine learning methods [22,37].…”

Task Runtime Prediction in Scientific Workflows Using an Online Incremental Learning Approach

“…To the best of our knowledge, this is the first work that predicts the runtime of workflow tasks using an online incremental learning approach. Hence, to compare our work with existing state-of-the-art solutions of task runtime prediction, we reproduce the batch offline learning work by da Silva et al [9] that makes use of a task's input data as a feature to predict the task runtime. We refer to this approach as the baseline scenario.…”