DOKK / manpages / debian 12 / bpfcc-tools / dirtop-bpfcc.8.en
dirtop(8) System Manager's Manual dirtop(8)

dirtop - File reads and writes by directory. Top for directories.

dirtop -d directory1,directory2,... [-h] [-C] [-r MAXROWS] [-s {reads,writes,rbytes,wbytes}] [-p PID] [interval] [count]

This is top for directories.

This traces file reads and writes, and prints a per-directory summary every interval (by default, 1 second). By default the summary is sorted on the highest read throughput (Kbytes). Sorting order can be changed via -s option.

This uses in-kernel eBPF maps to store per process summaries for efficiency.

This script works by tracing the __vfs_read() and __vfs_write() functions using kernel dynamic tracing, which instruments explicit read and write calls. If files are read or written using another means (eg, via mmap()), then they will not be visible using this tool. Also, this tool will need updating to match any code changes to those vfs functions.

This should be useful for file system workload characterization when analyzing the performance of applications.

Note that tracing VFS level reads and writes can be a frequent activity, and this tool can begin to cost measurable overhead at high I/O rates.

Since this uses BPF, only the root user can use this tool.

CONFIG_BPF and bcc.

Defines a list of directories, comma separated, to observe. Wildcards are allowed if between single bracket.
Don't clear the screen.
Maximum number of rows to print. Default is 20.
Sort column. Default is rbytes (read throughput).
Trace this PID only.
Interval between updates, seconds.
Number of interval summaries.

# dirtop -d '/hdfs/uuid/*/yarn'
# dirtop -d '/hdfs/uuid/*/yarn' -Cr 8
5 second summaries, 10 times only:
# dirtop -d '/hdfs/uuid/*/yarn' 5 10
# dirtop -d '/hdfs/uuid/*/yarn,/hdfs/uuid/*/data'

The contents of /proc/loadavg
Count of reads during interval.
Count of writes during interval.
Total read Kbytes during interval.
Total write Kbytes during interval.
The path were the IOs were accounted.

Depending on the frequency of application reads and writes, overhead can become significant, in the worst case slowing applications by over 50%. Hopefully for real world workloads the overhead is much less -- test before use. The reason for the high overhead is that VFS reads and writes can be a frequent event, and despite the eBPF overhead being very small per event, if you multiply this small overhead by a million events per second, it becomes a million times worse. Literally. You can gauge the number of reads and writes using the vfsstat(8) tool, also from bcc.

This is from bcc.

https://github.com/iovisor/bcc

Also look in the bcc distribution for a companion _examples.txt file containing example usage, output, and commentary for this tool.

Linux

Unstable - in development.

Erwan Velu

filetop(8) by Brendan Gregg

vfsstat(8), vfscount(8), fileslower(8)

2020-03-16 USER COMMANDS