Quantifying the Overheads of the Modern Linux I/O Stack

Authors: Luke Logan, Anthony Kougkas, and Xian-He Sun (Illinois Institute of Technology)

Abstract: The performance of the Linux I/O stack is critical to the performance of distributed storage applications. Recent research has shown that the Linux I/O stack introduces multiple overheads that significantly reduce and randomize the performance of I/O operations. A lesser amount of research has been conducted into quantifying the software overheads in the Linux I/O stack. In this research, we quantify this overhead by tracing the POSIX read()/write() system calls on various storage devices and filesystems. By comparing the amount of time spent in software versus the amount of time spent in performing I/O, we can gain insight on how much overhead the Linux I/O stack produces and explore solutions that can mitigate these overheads.

We will be presenting the amount of time spent in different layers of the Linux I/O stack to show where the overheads are, and we will explore potential ways to bypass these overheads.

Best Poster Finalist (BP): no

Poster: PDF
Poster summary: PDF


Back to Poster Archive Listing