Improving Operating System Support for Concurrent I/O

Colloq: Speaker: 
Chuanpeng Li
Colloq: Speaker Institution: 
University of Rochester
Colloq: Date and Time: 
Thu, 2008-03-06 10:00
Colloq: Location: 
5700, L202
Colloq: Host: 
Phil Roth
Colloq: Host Email: 
rothpc@ornl.gov
Colloq: Abstract: 
Modern computing systems become increasingly data-driven. Concurrent I/O is a commonplace in data-intensive server systems and parallel computing systems. In this talk, I will introduce two techniques we proposed for improving OS support for concurrent I/O. First, during concurrent I/O workloads, frequent disk head switching between multiple I/O streams may severely affect I/O efficiency. Aggressive prefetching can improve the granularity of sequential data access, but it comes with a higher risk of fetching unneeded data. We propose a competitive prefetching strategy that controls prefetching depth so that the overhead of disk I/O switch and unnecessary prefetching are balanced. Second, concurrent I/O systems may contain a large amount of prefetched data in memory. Traditional access-history based page reclamation methods are not effective for prefetched pages. We propose a heuristic-based OS-level scheme that can manage prefetch memory more efficiently.<br><br>Beyond this work, I will also briefly introduce some of my other research efforts with my colleagues, including I/O system performance debugging and operating system exploitation of processor hardware event counters.
Colloq: Speaker Bio: 
N/A