Author

Jeremy Logan

Date of Award

2010

Level of Access Assigned by Author

Open-Access Thesis

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

Advisor

Phillip Dickens

Second Committee Member

George Markowsky

Third Committee Member

James Fastook

Abstract

Today's most advanced scientific applications run on large clusters consisting of hundreds of thousands of processing cores, access state of the art parallel file systems that allow files to be distributed across hundreds of storage targets, and utilize advanced interconnections systems that allow for theoretical I/O bandwidth of hundreds of gigabytes per second. Despite these advanced technologies, these applications often fail to obtain a reasonable proportion of available I/O bandwidth. The reasons for the poor performance of application I/O include the noncontiguous I/O access patterns used for scientific computing, contention due to false sharing, and the somewhat finicky nature of parallel file system performance. We argue that a more fundamental cause of this problem is the legacy view of a file as a linear sequence of bytes. To address these issues, we introduce a novel approach for parallel I/O called Interval I/O. Interval I/O is an innovative approach that uses application access patterns to partition a file into a series of intervals, which are used as the fundamental unit for subsequent I/O operations. Use of this approach provides superior performance for the noncontiguous access patterns which are frequently used by scientific applications. In addition, the approach reduces false contention and the unnecessary serialization it causes. Interval I/O also significantly increases the performance of atomic mode operations. Finally, the Interval I/O approach includes a technique for supporting parallel I/O for cooperating applications. We provide a prototype implementation of our Interval I/O system and use it to demonstrate performance improvements of as much as 1000% compared to ROMIO when using Interval I/O with several common benchmarks.

Files over 10MB may be slow to open. For best results, right-click and select "save as..."

Share