Sequential File Prefetching in Linux

Abstract

Sequential prefetching is a well established technique for improving I/O performance. As Linux runs an increasing variety of workloads, its in-kernel prefetching algorithm has been challenged by many unexpected and subtle problems; As computer hardware evolves, the design goals should also be adapted. To meet the new challenges and demands, a prefetching algorithm that is aggressive yet safe, flexible yet simple, scalable yet efficient is desired. In this chapter, the author explores the principles of I/O prefetching and present a demand readahead algorithm for Linux. He demonstrates how it handles common readahead issues by a host of case studies. Both static, logic and dynamic behaviors of the readahead algorithm are covered, so as to help readers building both theoretical and practical views of sequential prefetching.

Omnisc'IO: A Grammar-Based Approach to Spatial and Temporal I/O Patterns Prediction

“…When our algorithm makes several predictions for the next context symbol (and therefore several predictions of offset), correct predictions are weighed accordingly. We compare our solution with the classical contiguous access estimation [17], which consists of always predicting that the next offset will follow the previous access. Table VI shows the proportion of contiguous accesses in our set of applications as well as the proportion of correct predictions made by Omnisc'IO.…”

“…Prefetching and caching indeed require a prediction of the location of future accesses [17], while I/O scheduling leverages estimations of I/O requests' interarrival time. Although these domains have been investigated for decades in the context of commodity computers [28], we restrict our study of related works mostly to their use in the HPC area, where applications have different (mostly more regular) I/O behavior.…”

“…b) Tracking offsets: Many prefetching systems, including those implemented in the Linux kernel [17], are based on the assumptions of consecutive accesses; that is, the next operation is likely to start from the offset where the previous one ended. As we will show in our experiments, this assumption is held for the POSIX-based applications that we tested, but it fails for applications that use a higher-level library such as HDF5.…”

Using Formal Grammars to Predict I/O Behaviors in HPC: The Omnisc'IO Approach

“…Many prefetching systems, including those implemented in the Linux kernel [18], are based on the assumptions of consecutive accesses; that is, the next operation is likely to start from the offset where the previous one ended. As we will show in our experiments, this assumption is held for the POSIX-based applications that we tested, but it fails for applications that use a higher-level library such as HDF5.…”

“…When our algorithm makes several predictions for the next context symbol (and therefore several predictions of offset), correct predictions are weighed according to our weighting system. We compare our solution with the classical contiguous access estimation [18], which consists of always predicting that the next offset will follow the previous access. Table 7 shows the proportion of contiguous accesses in our set of applications as well as the proportion of correct predictions made by Omnisc'IO.…”