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

offcputime - Summarize off-CPU time by kernel stack trace. Uses Linux eBPF/bcc.

offcputime [-h] [-p PID | -t TID | -u | -k] [-U | -K] [-d] [-f] [--stack-storage-size STACK_STORAGE_SIZE] [-m MIN_BLOCK_TIME] [-M MAX_BLOCK_TIME] [--state STATE] [duration]

This program shows stack traces and task names that were blocked and "off-CPU", and the total duration they were not running: their "off-CPU time". It works by tracing when threads block and when they return to CPU, measuring both the time they were off-CPU and the blocked stack trace and the task name. This data is summarized in the kernel using an eBPF map, and by summing the off-CPU time by unique stack trace and task name.

The output summary will help you identify reasons why threads were blocking, and quantify the time they were off-CPU. This spans all types of blocking activity: disk I/O, network I/O, locks, page faults, involuntary context switches, etc.

This is complementary to CPU profiling (e.g., CPU flame graphs) which shows the time spent on-CPU. This shows the time spent off-CPU, and the output, especially the -f format, can be used to generate an "off-CPU time flame graph".

See http://www.brendangregg.com/FlameGraphs/offcpuflamegraphs.html

This tool only works on Linux 4.6+. It uses the new `BPF_STACK_TRACE` table APIs to generate the in-kernel stack traces. For kernels older than 4.6, see the version under tools/old.

Note: this tool only traces off-CPU times that began and ended while tracing.

CONFIG_BPF and bcc.

Print usage message.
Trace this process ID only (filtered in-kernel).
Trace this thread ID only (filtered in-kernel).
Only trace user threads (no kernel threads).
Only trace kernel threads (no user threads).
Show stacks from user space only (no kernel space stacks).
Show stacks from kernel space only (no user space stacks).
Insert delimiter between kernel/user stacks.
Print output in folded stack format.
Change the number of unique stack traces that can be stored and displayed.
The minimum time in microseconds over which we store traces (default 1)
The maximum time in microseconds under which we store traces (default U64_MAX)
Filter on this thread state bitmask (eg, 2 == TASK_UNINTERRUPTIBLE). See include/linux/sched.h for states.
Duration to trace, in seconds.

# offcputime
# offcputime 5
# offcputime -f 5
# offcputime -p 185

This summarizes unique stack traces in-kernel for efficiency, allowing it to trace a higher rate of events than methods that post-process in user space. The stack trace and time data is only copied to user space once, when the output is printed. While these techniques greatly lower overhead, scheduler events are still a high frequency event, as they can exceed 1 million events per second, and so caution should still be used. Test before production use.

If the overhead is still a problem, take a look at the MINBLOCK_US tunable in the code. If your aim is to chase down longer blocking events, then this could be increased to filter shorter blocking events, further lowering overhead.

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.

Brendan Gregg

stackcount(8)

2016-01-14 USER COMMANDS