This paper, on the autonomic management of application workflows on a hybrid computing infrastructure, is designed to address the system and application heterogeneity and dynamics, as well as to ensure that the objectives and constraints of applications are satisfied--specifically, to demonstrate how different application objectives can be achieved while satisfying deadline and budget constraints using an appropriate mix of resources from the public and private cloud structure. Three types of objectives are considered: acceleration of application time-to-completion, using hybrid resources; conservation of high-performance computing (HPC) resources by limiting the usage of central processing unit (CPU) cores or CPU time when using public clouds; and, finally, resilience during some resource failure and unexpected delays. The hybrid and dynamic computing infrastructure described in this paper can be divided into two aspects: resource provisioning and adaptation.

The first aspect--the scheduling and provisioning of each task--is done by mapping the resource class that best accommodates the user objectives described above. Tasks are mapped, resources are allocated, and tasks are monitored. The autonomic manager and the adaptivity manager work together to ensure the objectives are met. The autonomic manager responsible for task scheduling has four components: workflow manager, runtime estimator, autonomic scheduler, and grids/clouds/clusters agents. The workflow manager coordinates the execution of the overall tasks, based on the user objectives and constraints mentioned; the runtime estimator computes the runtime and the cost of each task; the automatic scheduler uses the information from the estimator to determine the optimal initial mix of resources based on the constraints and objectives; and the grids and clouds are responsible for allocating the “resources on their specific platforms [and for] configuring workers as execution agents on these resources.” The adaptivity manager is responsible for observing the task runtimes, which workers report to the master in the results. Thus, the analyzer re-estimates the runtime of remaining tasks based on the historical data gathered from the results. Finally, the adapter receives a request for rescheduling and in turn calls the autonomic manager to schedule and retrieve the remaining tasks accordingly.

The experiment was conducted using a hybrid infrastructure with TeraGrid and five instance types of Amazon EC2. The number of cores on TeraGrid was set to 16. The results of acceleration using greedy deadline, economical deadline, and budget limitations were presented. This paper mainly shows that public clouds can be used to achieve user performance and cost objectives, and recover from unexpected delays and failure. The authors could have explained in more detail how the tasks are mapped to different nodes in this autonomic management system. There was one simple example of task mapping, but the topic could have been explained in better detail with more complex ones.

The paper proposes a whole new autonomic management system based on resource provisioning and resource adaptation. I have some lingering questions: Are all of the drawbacks presented? Is there any situation where the constraints are not met? It is an important issue to address the users’ constraints based on deadline and budget. Are there more constraints that need to be included to satisfy the users’ needs? Perhaps these questions will be addressed in a future paper.