SC Conference - Activity Details

Feedback-controlled, Virtualized Resource Sharing for Predictable eScience

Sang-Min Park  (University of Virginia)
Doctoral Research Showcase Session
Thursday,  04:15PM - 04:30PM
Room 17A/17B
The emerging dynamic data driven applications are key to the time-critical discoveries such as tornado warnings. While the applications require timely execution to process sporadic sensor events in real-time, the prevailing batch-mode job processing cannot provide a sufficient level of predictability. The widely-discussed reservation approach only partially solves the problem yet it creates severe resource underutilization. Predictable application execution has the two key requirements: the ability to change completion time of jobs at fie-granularity; and the ability to cope with unanticipated disturbances that affect a time-dependent job's performance We build Compute Throttling Framework to address the two requirements. The first requirement is solved via a novel use of virtualization technology that enables fine-grained, online resource control. We use formal control theory to solve the second requirement. Control theory allows us to design an adaptive resource controller that regulates job's progress and copes with unforeseen disturbances in order to meet the job's deadline.
   IEEE Computer Society  /  ACM     2 0   Y E A R S   -   U N L E A S H I N G   T H E   P O W E R   O F   H P C