IMPROVING PAGING PERFORMANCE OF MEMORY-INTENSIVE APPLICATIONS WITH MEMORY ACCESS PATTERN GUIDED PAGE REPLACEMENT
Nawab Ali · OhioLink ETD Center (Ohio Library and Information Network) · 2004
Memory-intensive applications have large data sets which cannot be accommodated in the available main memory.The Virtual Memory (VM) sub-system frequently moves parts of their data sets between the main memory and the secondary storage (swap disk).This results in a significant increase in the paging I/O traffic of the applications.Since the disk sub-system is orders of magnitude slower than the main memory, this increase in I/O traffic severely degrades the applications' performance.In this research, we reduce the paging I/O traffic of memory-intensive applications by exploiting the patterns in their memory access behavior.Most out-of-core applications exhibit clear patterns in the manner in which they access their virtual address space.We collect and analyze the memory access patterns of these applications and use this information to guide the page replacement algorithm.This technique, when coupled with memory prefetching and page clustering results in a substantial performance improvement for large, memory-intensive applications.We conducted trace-driven simulations on the NAS Parallel Benchmarks (NPB) to evaluate our paging optimization algorithms.The simulation results show a considerable improvement in the performance of our benchmarks.Our page replacement strategy reduced the average paging I/O traffic by 52%.On combining memory prefetching with our pattern guided page replacement algorithm, we observed an average performance improvement of up to 69%.These results are encouraging, and we hope to further improve the VM performance of memory-intensive applications by adding hardware support to our strategy.This research would not have been possible without the assistance, patience and support of many individuals.I would like to extend my gratitude to my thesis advisor Dr. Yiming Hu for mentoring me over the course of my graduate studies.His guidance, insight and encouragement helped me overcome the numerous challenges I faced during this research.I would also like to thank Dr. Dharma Agrawal and Dr. Qing-An Zeng for serving on my thesis committee.I would additionally like to thank Rui Min for his work on the virtual memory performance of memory-intensive applications.Rui was extremely helpful in explaining his ideas and encouraged me to explore new research areas.