Jon Weissman, Principal Investigator

The iScheduler: A Synergistic Approach to Scheduling Computational Science Applications

Integrated scheduling in a grid.

Efficiently executing resource-intensive parallel scientific applications in dynamic parallel and distributed production environments is a challenging problem. Applications can experience large slowdowns or high wait times due to contention for these limited valuable resources especially as the scale of the application grows to realistic sizes. While the performance problems are rooted in the oversubscription of system resources, they are further exacerbated by inefficient resource utilization both by the system and the application. The problem is that application schedulers typically select resources without regard for the negative impact this may have upon other applications and may even over-allocate resources with limited benefit. In contrast, most job schedulers in parallel systems are often designed to optimize metrics that may not deliver the best possible application performance, but offer good average performance or favor machine utilization.

These researchers are investigating a new paradigm that narrows the functional gap between job and application scheduling called integrated scheduling or iScheduling. The goal of iScheduling is to exploit the advantages of job scheduling (global view of competing applications) with application scheduling (specific knowledge of an application) to yield reduced time-to-solution for computational science applications and make more efficient use of available hardware for high performance systems. Within the iScheduler, resource allocation decisions are more fluid than in most static scheduling systems. To achieve this, the iScheduler and application interact using an interface that provides dynamic application performance information to the iScheduler (sensing) and an interface that is used to control the application’s behavior and resource usage in response to information gathered by sensing (actuating). The researchers are establishing the efficacy of iSchedulers by building several instances and evaluating their performance on computer clusters.

Research Group

Lakshman Abburri Rao, Graduate Student Researcher