Hiding the latency of paging and input/output operations on massively parallel systems
Abstract
In this thesis, we studied the behavior of parallel programs to understand how to automated the task of hiding latency of paging, input/output, and communication operations on massively parallel processing systems. We designed a parallel performance monitoring environment, the Musketeers, and investigated its use to improve the performance of parallel programs on DMIMD systems. In designing the Musketeers, we examined the interference of the instrumentation on the execution of parallel programs and presented some alternatives to minimize its effects. Since collecting performance data by monitoring program execution is only the first step to understanding the behavior of programs, we provided several customized monitoring environments and analysis methods for studying paging, I/O, and communication activity of parallel programs. We devised the skyline analysis method to correlate the paging activity of processes running concurrently regardless of the size of the process group. Our cumulative analysis gives information about the loads placed on system resources. Using simple analysis methods our studies revealed several I/O performance bottlenecks for actual parallel applications. In investigating communication events, in particular the relationship between synchronization requirements and scheduling of process groups, we found that temporal locality of communication can be used by a scheduler to improve the overall utilization of system resources.
Degree
Ph.D.
Advisors
Marinesch, Purdue University.
Subject Area
Computer science
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.