Check out the new USENIX Web site.

Home About USENIX Events Membership Publications Students
Abstract - Technical Program - OSDI 99

Automatic I/O Hint Generation Through Speculative Execution

Fay Chang, Garth A. Gibson
Carnegie Mellon University


Aggressive prefetching is an effective technique for reducing the execution times of disk-bound applications; that is, applications that manipulate data too large or too infrequently used to be found in file or disk caches. While automatic prefetching approaches based on static analysis or historical access patterns are effective for some workloads, they are not as effective as manually-driven (programmer-inserted) prefetching for applications with irregular or input-dependent access patterns. In this paper, we propose to exploit whatever processor cycles are left idle while an application is stalled on I/O by using these cycles to dynamically analyze the application and predict its future I/O accesses. Our approach is to speculatively pre-execute the application's code in order to discover and issue hints for its future read accesses. Coupled with an aggressive hint-driven prefetching system, this automatic approach could be applied to arbitrary applications, and should be particularly effective for those with irregular and, up to a point, input-dependent access patterns.

We have designed and implemented a binary modification tool, called "SpecHint", that transforms Digital UNIX application binaries to perform speculative execution and issue hints. TIP [Patterson95], an informed prefetching and caching manager, takes advantage of these application-generated hints to better use the file cache and I/O resources. We evaluate our design and implementation with three real-world, disk-bound applications from the TIP benchmark suite. While our techniques are currently unsophisticated, they perform surprisingly well. Without any manual modifications, we achieve 29%, 69% and 70% reductions in execution time when the data files are striped over four disks, improving performance by the same amount as manually-hinted prefetching for two of our three applications. We examine the performance of our design in a variety of configurations, explaining the circumstances under which it falls short of that achieved when applications were manually modified to issue hints. Through simulation, we also estimate how the performance of our design will be affected by the widening gap between processor and disk speeds.

  • View the full text of this paper in HTML form and PDF form.

  • If you need the latest Adobe Acrobat Reader, you can download it from Adobe's site.

  • To become a USENIX Member, please see our Membership Information.

?Need help? Use our Contacts page.

Last changed: 27 Mar 2002 ml
Technical Program
Symposium Index