2014-04-27 19:28:58

by Mason

[permalink] [raw]
Subject: Using ftrace to identify source of excessive latency of USB write

Hello everyone,

I'm using Linux on a embedded system similar in spec to a desktop PC
from 15 years ago (256 MB RAM, 800-MHz CPU, USB). The system's primary
use is recording high-definition digital television programs.

Typically, the storage sub-system consists of a recent hard-disk drive
connected over USB (Hi-Speed, effective throughput ~20-30 MB/s), using
a single ext4 partition (journal disabled), mounted noexec+noatime
(trying to minimize metadata interference).

Typical bit-rate for this HD content ~1-3 MB/s

Data is accumulated in two 800-kB buffers; when one buffer is full,
it is written to file (using write(2)), which was opened O_SYNC.
(Note to self: try O_DSYNC instead of O_SYNC)

If I plot the latency of the write(2) operation, 99% of them complete
in under 80 ms. However, in rare cases, there is a huge latency spike
(up to 800 ms). If several of these rare outliers occur in a row,
the recording is messed up.

I am trying to figure out the source of these latency spikes.

It could be the OS, the USB controller, the HDD controller... I was
hoping I could use ftrace to determine whether the problem came from
the OS itself. Is that the best tool for the job?

Any recommendations on how to proceed?

Regards.

[I would be grateful if you could CC me in your replies. Thanks!]