Time and cost-efficient resource allocation for real-time application in high performance computing systems

Abstract

High Performance Computing (HPC) is the de-facto platform for deploying real-time applications due to the collaboration of large-scale resources operating in cross-administrative domains. HPC resource scheduling and allocation is a crucial issue in achieving efficient utilization of available resources, especially when resource-intensive applications have real-time deadlines and need data files replicated over the data storage resources. Such scheduling engages both computing and data storage resources to carry out application execution in a timely manner. Traditional approaches are sufficient only when data storage resources are coupled with the computing resources in HPC environment, since data is available at the computing resources for application execution. However, the said domain leaves gaps for deadline miss when data is transferred from remotely located data storage resources to the computing resources where application is being executed. The deadline miss mainly occurs due to the unavailability of the required data files, inadequate scheduling and allocation mechanism of the HPC resources. The problem becomes more complicated when some of the data files are pre-fetched while some post-fetched during application execution which usually results in delayed processing and in turn deadlines miss. The allocation of such resources by considering different optimization criteria such as makespan minimization, cost and energy efficiency, respecting application deadlines, etc. in the aforementioned scenario can be gracefully addressed by designing a scheduling strategy which can result in improved resources utilization while predicting application feasibility. It has always been of interest to the research community to pose the abovementioned situation to determine if the existing scheduling theory and resource allocation strategies are mature enough to accommodate the challenges presented with the emergence of the latest HPC platforms. In this thesis, we explore and analyze the existing resource-allocation techniques for scheduling real-time applications with temporal constraints on HPC platforms (grid, cloud, edge, fog, and multicore systems). This study further compares the resource allocation mechanisms based on different performance parameters and based on existing gaps, a model is proposed which predicts the application schedulability by analyzing time and data constraints before actually dispatching the application to the HPC resources. The main advantage of the prediction-based model is to save time by declining further analysis on unsuitable resources which improve resource utilization by considering application workload in advance. Furthermore, this research thesis devises time and cost-efficient variants of HPC resource allocation with provably correct formulations to cope with the aforementioned problems so that both the user and real-time application constraints are respected. The most celebrated results affirm the supremacy of the proposed techniques in obtaining the desired level of service.