Performance Management of Enterprise Application Systems in the Cloud Era

企业应用，如Web和交互式大数据服务，需要快速响应用户事务。因此，系统操作员需要确保应用程序满足其响应时间目标的技术，同时以经济有效的方式利用计算资源。有几个因素需要为这样的系统提供新的性能管理技术。例如，这些应用程序越来越多地部署在公共云平台上，由于共享云资源的争用，这些平台可能会出现不可预测的性能下降。需要新的方法来管理存在这种平台诱导干扰的系统性能。此外，这些系统通常会遇到突发工作负载，这会以复杂的方式降低性能。这激发了能够预测和减轻爆炸影响的新技术。本项目旨在解决这些挑战。我们将研究新技术，使运营商能够准确预测系统所需的云资源，以在处理给定工作负载时满足所需的响应时间目标。基于排队分析的技术通常需要专家手动创建系统模型。此外，在预测突发工作负载时，准确性也会受到影响。机器学习（ML）技术有望成为数据驱动的排队分析替代方案。然而，现有的工作并没有提供对准确性有很大影响的任务的清晰直觉，比如ML技术选择、特征化和训练数据选择。我的程序将解决这一知识差距，并实现自动预测技术，不负担操作员这样的任务。我们还将探索运行时技术，以减轻突发和干扰的影响。现有的工作没有集中于处理服务需求突发的不利影响，即导致系统资源持续高利用率或低利用率的用户事务模式。我们的初步工作表明，这种突发事件可以通过智能地重新排序传入的事务来控制，只需使用更少的资源。我们将在此基础上实现新的运行时调度技术。作为本主题的一部分，我们还将利用我们正在进行的干扰检测工作来自动扩展云资源实例，例如容器，以响应干扰。现有的方法没有考虑单个事务类型如何受到给定实例中干扰的影响。我们将构建可以使用这种细粒度信息来智能地将事务分发到实例的模型，从而使用最小的实例来减轻干扰。该计划将扩大在数据驱动的性能预测和管理研究的艺术状态。加拿大的组织可以利用这项研究来降低与性能差和资源供应过剩相关的成本。

Enterprise applications, e.g., Web and interactive big data services, need to respond quickly to user transactions. Consequently, system operators need techniques that ensure applications meet their response time objectives while utilizing computing resources in a cost-effective way. Several factors necessitate new performance management techniques for such systems. For example, these applications are being increasingly deployed on public cloud platforms, which can suffer from unpredictable performance degradations due to contention for shared cloud resources. Novel approaches are needed to manage system performance in the presence of such platform induced interference. Furthermore, these systems typically experience bursty workloads, which can degrade performance in complex ways. This motivates new techniques that can predict and mitigate the impact of burstiness. This program seeks to address such challenges. We will investigate new techniques that allow an operator to accurately predict the cloud resources needed by a system to satisfy a desired response time target while handling a given workload. Techniques based on queuing analysis typically require an expert to manually author a system model. Also, accuracy can be impacted when predicting for bursty workloads. Machine learning (ML) techniques promise a data-driven alternative to queuing analysis. However, existing work does not provide clear intuition on tasks that can have a big impact on accuracy such as ML technique selection, featurization, and training data selection. My program will address this knowledge gap and realize automated prediction techniques that do not burden an operator with such tasks. We will also explore runtime techniques to mitigate the impact of burstiness and interference. Existing work has not focused on handling the adverse impact of service demand burstiness, i.e., user transaction patterns that cause sustained periods of high or low utilizations at system resources. Our initial work suggests that such burstiness can be tamed using fewer resources by intelligently reordering incoming transactions. We will build on this insight to realize new runtime scheduling techniques. As part of this theme, we will also exploit our ongoing work on interference detection to automatically scale cloud resource instances , e.g., containers, in response to interference. Existing approaches do not consider how individual transaction types get impacted by interference at a given instance. We will build models that can use such fine-grained information to intelligently distribute transactions to instances such that interference is mitigated using minimum instances.This program will expand the state of the art in data-driven performance prediction and management research. Canadian organizations can exploit the research to reduce costs related to poor performance and resource over-provisioning.