Cues: File Subsystem Data Streams

Abstract

We present a system call which enable directing high performance data transfers with in-kernel streams. The streams are defined and run using abstract data types called Cues, formed over the Unix file subsystem. The system call, named cue() , returns a descriptor, which can be used to write requests to the newly created stream. These requests define the flow of the stream. The abstraction simplifies the design of applications that transfer large amounts of data from files or devices. It also enables high throughput when multiple transfers are in progress. Cue code is compact, modular, and portable. This model also results in a simple mechanism for remotely cueing data flow using standard connections with peer processes. The implementation and tests are also described in this paper.