PIOSS: A Simulation Model for the Analysis of Parallel I/O Performance Variability on Large-scale Applications

Abstract

To meet ever increasing capacity and performance requirements of emerging data-intensive applications, parallel file systems (PFSs) have been employed in large-scale computing environments. In such complex storage systems, the load distribution on PFS data servers compose a major source of input/output (I/O) performance variability. Albeit mitigating such variability is desirable, understanding its sources and behavior remains a challenging task. In this research work, a differentiated approach for evaluating the parallel I/O performance variability perceived by large-scale applications is proposed. The Parallel I/O and Storage System (PIOSS) simulation model represents main components and mechanisms observed in typical PFS implementations and enables fast evaluations of large and complex scenarios. Experimental results presented in this paper demonstrate PIOSS can accurately reproduce the load balance on PFS data servers, with a confidence level of 95%.