Re: Benchmarks of kernel tracing options (ftrace, ktrace, lttng and perf)

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

* David Sharp (dhsharp@google.com) wrote:
[...]
> Results for amount of time to execute a tracepoint (includes previous results):
> ktrace: 200ns   (old)
> ftrace: 224ns   (old, w/ handcoded tracepoint, not syscall tracing)
> lttng: 449ns    (new)
> perf: 1047ns    (new)
> 
> Also interesting:
> ftrace: 587ns   (old, w/ syscall tracing)
> This just shows that syscall tracing is much slower than a normal tracepoint.

As I pointed out in my email a few weeks ago, the LTTng comparison is simply
bogus because the "syscall tracing" thread-flag is active, which calls into
syscall tracing, after saving all registers, from entry_*.S, both at syscall
entry and exit.

I did benchmarks using Steven's ring_buffer_benchmark kernel module, which calls
tracing in a loop, for both Ftrace and the Generic Ring Buffer (which is derived
from LTTng). The results are:

Intel Xeon 2.0GHz

ftrace:  103 ns/entry (no reader)
lttng:    83 ns/entry (no reader) (with the generic ring buffer library)

So, given that even after I pointed out that the results above were bogus,
people took the numbers for granted, and given that David seems busy on other
things, I thought I should set records straight.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com