Demonstrations of disksnoop.py, the Linux eBPF/bcc version.


This traces block I/O, a prints a line to summarize each I/O completed:

# ./disksnoop.py 
TIME(s)            T  BYTES    LAT(ms)
16458043.435457    W  4096        2.73
16458043.435981    W  4096        3.24
16458043.436012    W  4096        3.13
16458043.437326    W  4096        4.44
16458044.126545    R  4096       42.82
16458044.129872    R  4096        3.24
16458044.130705    R  4096        0.73
16458044.142813    R  4096       12.01
16458044.147302    R  4096        4.33
16458044.148117    R  4096        0.71
16458044.148950    R  4096        0.70
16458044.164332    R  4096       15.29
16458044.168003    R  4096        3.58
16458044.171676    R  4096        3.59
16458044.172453    R  4096        0.72
16458044.173213    R  4096        0.71
16458044.173989    R  4096        0.72
16458044.174739    R  4096        0.70
16458044.190334    R  4096       15.52
16458044.196608    R  4096        6.17
16458044.203091    R  4096        6.35

The output includes a basic timestamp (in seconds), the type of I/O (W == write,
R == read, M == metadata), the size of the I/O in bytes, and the latency (or
duration) of the I/O in milliseconds.

The latency is measured from I/O request to the device, to the device
completion. This excludes latency spent queued in the OS.

Most of the I/O in this example were 0.7 and 4 milliseconds in duration. There
was an outlier of 42.82 milliseconds, a read which followed many writes (the
high latency may have been caused by the writes still being serviced on the
storage device).