FIO Tracer

An I/O tracing utility based on iobs which utilizes fio and blktrace to track the latencies of an IO across the Linux Kernel. This requires an additional 'S' tracepoint to be added, similar to what's done here. The following Linux tools are utilized internally: blktrace and fio.

Installation

The latest version can be obtained via wget:

$ wget https://raw.githubusercontent.com/uofl-csl/fio-tracer/master/fio-tracer.py

Getting Started

Executing the script without any arguments shows the different arguments that can be used:

$ sudo python3 fio-tracer.py
fio-tracer.py 0.1.0
Usage: fio-tracer.py <file> [-c] [-l] [-o <output>] [-r <retry>] [-v] [-x]
Command Line Arguments:
<file>            : The configuration file to use.
-o <file          : (OPTIONAL) Output metrics to a csv file.

Input

The main input to the tool should be a file in INI format which has the configurations for how to run each of the workloads. The possible configuration settings for each workload is the following:

• fio_command [str]- The workload generation command to run (e.x. fio ...).
• fio_file [str] - The file fio will operate on.
• fio_lat_file_pref [str] - The latency file prefix that will hold the latencies generated by FIO.
• fio_ramp_time [str] - If ramp time is used, indicate here to skip IOs prior to this.
• blktrace_delay [int] - The amount of delay between running the workload and starting the trace. Defaults to 0.
• device [str] - The device to run the trace on (e.x. /dev/sdd).
• schedulers [str] - The io schedulers to use (e.x. noop, cfq, deadline, none, mq-deadline, bfq, kyber).
• runtime [int] - The amount of time in seconds to run the trace (should match the workload runtime).
• filter_stddev [float] - If given, filters IOs with latencies > stddev * filter_stddev.
• filter_abnormal [float] - If given, filters IOs which have a negative latency in the file system layer.

Each of these commands should be under a header [...] indicating a specific job. The global header [global] can be used for configuring settings that are the same for all jobs.

Output

Upon completion of each job, the following metrics are output to STDOUT:

• CLAT (Completion Latency) [µs]
• SLAT (Submission Latency) [µs]
• Q2C [µs]
• Q2D (Block Layer Latency) [µs]
• D2S (Driver Latency) [µs]
• S2C (Device Latency) [µs]
• SLAT - Q2C (File System Latency) [µs]

The following is a sample output the is given for a job:

Results of workload:
clat [µs]: 30.80
slat [µs]: 8.02
q2c [µs]: 28.97
q2d [µs]: 10.35
d2c [µs]: 18.62
d2s [µs]: 2.30
s2c [µs]: 16.32
fs [µs]: 1.83

In addition to STDOUT, an output file can be given via the [-o] < output> command-line argument. This file will be written to as a csv with the first row being a header of the following columns:

• workload
• scheduler
• rw
• clat
• user
• block layer
• driver
• device

License

Copyright (c) 2018 UofL Computer Systems Lab. See LICENSE for details.