In the past decade, the gap between processing speeds and disk access times has increased by an order of magnitude . At the same time, although memory sizes have increased substantially, so have application data requirements. Systems therefore continue to swap their data sets to and from disk as they are often too large to all fit in memory.
In recognition of this problem, there has been a great deal of research into automating disk prefetching algorithms that are dramatically more accurate than the standard heuristics in current operating systems. Recent work by Demke and Mowry  demonstrated impressive performance results using compiler-based techniques. However, system-wide use of their approach would require recompiling every application. Moreover, their compiler analyses are limited to looping array codes.
To address this problem, Chang and Gibson  proposed to discover the future data accesses of an application stalled on I/O by executing ahead in its code, ignoring non-memory-resident data. They also proposed a design for automatically modifying application binaries to apply this speculative execution approach. With this design, they demonstrated that speculative execution can deliver substantial performance benefits for a diverse set of applications that issue explicit file read calls.
Unfortunately, their user-level design has several major disadvantages. First, their design relies on the correct implementation of a complex binary modification tool. Second, while their design does not require application source code, applications must still be transformed to receive benefit. Third, their design can only issue prefetches for disk reads caused by explicit file I/O because page faults cannot be trapped by the application. The system therefore cannot be applied to swapping (or out-of-core) applications which leverage file- and swap-backed virtual memory to avoid the complexity of explicit I/O. Finally, their design does not measure or limit the memory used by speculative execution, and its resulting effect on system performance. It is therefore poorly suited for use on realistic systems, which do not always have abundant memory.
In this paper, we demonstrate how these problems can be overcome with an in-kernel design for automating speculative execution. We present a design that, by leveraging existing operating system features, is not only substantially easier to implement but also may provide benefit to arbitrary unmodified application binaries. Moreover, by exploiting knowledge that is typically unavailable outside the operating system, our design automates disk prefetching for virtual memory accesses as well as explicit I/O calls, enabling it to provide benefit regardless of the I/O-access methods used by applications. In addition, we propose two new operating system features which substantially improve the performance of speculative execution for swapping applications. Finally, we describe a mechanism for estimating the impact of memory use by speculative executions on system performance, and thereby controlling speculative execution when memory resources are not abundant.
We have implemented our design within the Linux 2.4.8 kernel and evaluated it using eight applications, which include four explicit-I/O applications, three swapping applications, and one application that performs a substantial amount of both forms of I/O. We demonstrate that our basic design provides similar benefit to the prior design on explicit-I/O applications, while requiring much less implementation effort. We then demonstrate that our design also provides benefit for swapping applications, particularly with the assistance of the new features we identified. Finally, by varying the amount of usable memory on our system, we demonstrate the benefit of our mechanism for controlling the memory usage of speculative execution.
The remainder of this paper is organized as follows. Section 2 describes the speculative execution approach to automating I/O prefetching. Section 3 describes the testbed and benchmarks that we use to evaluate our proposals throughout this paper. Section 4 presents and evaluates the baseline version of our new in-kernel design and implementation. Then, in Sections 5 and 6, we proceed to describe our improvements for swapping applications, and our mechanism for controlling memory overhead. Section 7 contrasts our in-kernel design with the prior user-level design. Finally, Sections 8 and 9 discuss related work, future work, and conclusions.