ivring driver

ivring driver

[Mathieu Desnoyers]

Hi Yoshihiro,

I stumbled on your post on ivring (https://lkml.org/lkml/2012/6/5/143)
for buffering trace data from guest to host. I notice that it would be
very interesting to see a comparison of its performance against:

- ftrace ring buffer
- perf ring buffer
- lttng 2.0 ring buffer
- lttng-ust 2.0 (user-space tracing) ring buffer

Comparing only with network-related mechanisms does not seem to provide
a complete picture of where it stands wrt other ring buffers out there.

One of the metric we have used in the past is the number of nanoseconds
it takes to write a single event, based on a test that writes lots of
events into various buffer sizes, into flight recorder mode buffers
(which overwrites oldest information on buffer full condition, with data
collection to disk disabled, to remove I/O from the picture -- I/O can
then be performed in "snapshot" mode, after some error condition is
detected).

Best regards,

Mathieu

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

Re: ivring driver

[Yoshihiro YUNOMAE]

(2012/06/06 0:29), Mathieu Desnoyers wrote:
> Hi Yoshihiro,
>
> I stumbled on your post on ivring (https://lkml.org/lkml/2012/6/5/143)
> for buffering trace data from guest to host. I notice that it would be
> very interesting to see a comparison of its performance against:
>
> - ftrace ring buffer
> - perf ring buffer
> - lttng 2.0 ring buffer
> - lttng-ust 2.0 (user-space tracing) ring buffer

Hi Mathieu. Thank you for commenting on how to compare among
ring-buffer's performances. And sorry, I'm late for replying to you.

> Comparing only with network-related mechanisms does not seem to provide
> a complete picture of where it stands wrt other ring buffers out there.
>
> One of the metric we have used in the past is the number of nanoseconds
> it takes to write a single event, based on a test that writes lots of
> events into various buffer sizes, into flight recorder mode buffers
> (which overwrites oldest information on buffer full condition, with data
> collection to disk disabled, to remove I/O from the picture -- I/O can
> then be performed in "snapshot" mode, after some error condition is
> detected).

OK, I see. When the overhead of ring-buffers will be evaluated in
the next, I may use the metric as a reference.

Thank you,

-- 
Yoshihiro YUNOMAE
Software Platform Research Dept. Linux Technology Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: yoshihiro.yunomae.ez@hitachi.com