Fun with tracepoints [LWN.net]

Fun with tracepoints

Posted Aug 13, 2009 18:18 UTC (Thu) by karim (subscriber, #114)
In reply to: Fun with tracepoints by SEJeff
Parent article: Fun with tracepoints

Sorry, this has been tried and has failed. Check out the list of companies who have contributed to LTTng:
Google, IBM, Ericsson, Autodesk, Wind River, Fujitsu, Monta Vista, STMicroelectronics, C2 Microsystems, Sony, Siemens, Nokia.

But, hey, who are they to know, the kernel developers know better. And the hell with Sun, Apple, IBM, Microsoft, etc. who spent large gobs of money on implementing tracing infrastructure in their OSes (Apple by the way ported DTrace to MacOS ... :/ ) and maintaining it through the years. They're wrong too. The Linux kernel developers surely are better than the collective intelligence of the engineers and product managers of the aforementioned.

I forgot to mention that apart from pushing and maintaining LTT for a number of years, I also worked/defended a number of ideas which were dear to my heart. Take for example real-time. Very early on I came to the LKML pointing out that the tacit laissez-faire towards the RTLinux patent was not good for Linux. This was dismissed off-hand: the uses, I was told, were so narrow and the applications so specific that this is a non-issue ("real time apps are a niche market and they're not mainstream" ... i.e. those users don't matter). Skip a few years and there were two approaches being discussed Ingo's and the iPipe (my idea); at the subsequent OLS I asked a prominent developer whether what he thought were the chances of success of Ingo's very invasive approach, his reply was clear: Ingo has got the clout to make it happen. Just about then I knew iPipe wasn't likely to "win". And have his lunch he did.

That along with other things I witnessed (such as Con Kolivas quitting kernel development because he saw little interest in helping desktop interactivity) made me increasingly feel there's a NIH-syndrome. If nothing else, it distills from this that Linux' development has become highly politicized. You're either part of the in-crowd or you're not. And if you're not part of the in-crowd you're going to have a hell of a time trying to push something in if it's the least bit unconventional. Don't get wrong, being part of the in-crowd doesn't guarantee a radical change's acceptance. But being an outsider clearly ensures that you've got zero chances of success. It might have changed since I've stopped keeping track of it all, but juxtapose the previous with the fact that most kernel developers work for/on big iron and you've got a huge disconnect with the realities of real-life mainstream users. It's not that user preoccupations aren't eventually taken care or fixed (ex.: udev/sysfs/devfs), it's just that an absolute non-priority. And *that* is a serious issue. Last I checked, Linux has been flatlining in the end-user market for a very long time. If the diagnosis I'm making out of the symptoms I've witnessed is the least bit right (and I really hope I'm wrong), this isn't about to change any time soon.

I sincerely apologize if I've offended anyone with the above, but this is a case where *everything* ***EVERYTHING*** has been tried to convince the kernel development community. The ball is in their camp.

Fun with tracepoints

Posted Aug 18, 2009 21:40 UTC (Tue) by oak (guest, #2786) [Link] (6 responses)

> If nothing else, it distills from this that Linux' development has
> become highly politicized. You're either part of the in-crowd or
> you're not.

I think the problem is more that individual kernel developers don't really
(need to) look at the whole system or be responsible for it, just a one
corner of it. On commercial operating systems, there are dedicated people
who look after the whole thing and need to make sure that the whole thing
works fine (and this responsibility gives them influence over the
operating system implementation to make sure that these tools get done &
available).

If you're looking just at one or some parts of the whole system, things
like LTT (or to some extent Systemtap[1]) that try to get an overview of
what happens in the whole system may seem too large / complex /
intrusive / bloated. "I just need this specific info from the block
layer" (or memory subsystem, or ...). And then they write their own NIH
tracing for that single thing that doesn't much benefit others, or
somebody who wants to make sense out of the whole system.

Note: I have gotten useful info both from LTT and LTTng (lttng.org) + it's
finally getting easier to apply to kernel... LTTv plays a large part in
this too as one can easily zoom into details etc.

[1] Systemtap seems nice, but it doesn't have the post-processing /
visualization for the whole system like LTT does. I see it more like a
tool to do more specific analysis tools. However, for this kind of stuff
it's a bit too complicated (e.g. in embedded environments where you don't
want to run stap / compile the scripts on the device itself etc), so no
wonder devs write their own tracing...

Fun with tracepoints

Posted Aug 19, 2009 16:18 UTC (Wed) by fuhchee (guest, #40059) [Link] (5 responses)

[1] Systemtap seems nice, but it doesn't have the post-processing / visualization for the whole system like LTT does. I see it more like a tool to do more specific analysis tools.

I see what you mean. systemtap people are working on some GUI data graphing tools, but are just starting. (I got the impression though that LTTV was being deprecated in favour of eclipse-based widgets, which systemtap and other tools could feed data into also.)

However, for this kind of stuff it's a bit too complicated (e.g. in embedded environments where you don't want to run stap / compile the scripts on the device itself etc)

We hope to ease that pain by more automated cross-compilation/execution.

Fun with tracepoints

Posted Aug 20, 2009 19:29 UTC (Thu) by oak (guest, #2786) [Link] (4 responses)

> I got the impression though that LTTV was being deprecated in favour of
eclipse-based widgets

Do you have any pointers to more information about this?

Fun with tracepoints

Posted Aug 20, 2009 19:32 UTC (Thu) by fuhchee (guest, #40059) [Link] (3 responses)

I don't want to misrepresent LTTng, so please do take all this
with a grain of salt, but this is what I gathered from the presentations
given at http://ltt.polymtl.ca/tracingwiki/index.php/TracingMiniSu...

Fun with tracepoints

Posted Aug 20, 2009 19:34 UTC (Thu) by fuhchee (guest, #40059) [Link] (2 responses)

Grr, URL screwup, try this:
http://ltt.polymtl.ca/tracingwiki/index.php/TracingMiniSu...

Fun with tracepoints

Posted Aug 20, 2009 20:58 UTC (Thu) by oak (guest, #2786) [Link] (1 responses)

Thanks! According to this:
http://eclipse.org/linuxtools/projectPages/lttng/

"The first release, scheduled for September 2009 (code name: Vanilla),
will provide feature parity with the LTTng Viewer (LTTV) v0.12.12."

And this seemed to have a screenshot of the LTTng plugin:
http://ltt.polymtl.ca/tracingwiki/images/0/00/TMF_-_Traci...

This was pretty good overview of past & present tracing:
http://ltt.polymtl.ca/tracingwiki/images/5/57/Ts2009-hell...

Architecture bit here was annoying:
http://ltt.polymtl.ca/tracingwiki/images/4/46/Ts2009-Syst...

As it assumes that one has a working user-space in problematic cases one
wants to analyze. Kernel should be able to optionally get the data out
also through some high-speed HW interface without going through user-space
and filling of the flight record buffer would better not rely on
user-space.

Fun with tracepoints

Posted Aug 22, 2009 16:23 UTC (Sat) by compudj (subscriber, #43335) [Link]

About your comment on the architecture, I just want to clarify a few points. First, the architecture diagram you see at http://ltt.polymtl.ca/tracingwiki/images/4/46/Ts2009-Syst... focuses on tracing of userland. In this diagram, kernel tracing is contained within the "kernel trace facilities" box. For user-space tracing, where the goal is to get data out of the applications, it makes sense to consider than user-space is working. As you point out, this assumption makes less sense when we talk about tracing the kernel.

Second, more specifically about the LTTng kernel tracer, you are right in that the current mechanism used to extract data is a splice() system call controlled by a user-space daemon. However, alternate implementations of ltt-relay-alloc.c and ltt-relay-lockless.c could easily permit to use a high-speed debug interface. This has already been done with earlier LTTng versions for ARM.

The core of the LTTng kernel tracer therefore does not depend on userland. It's only the peripheral data extraction and trace control modules which depend on working userland. But they could be replaced easily by built-in kernel objects interacting directly with the LTTng kernel API. I made sure all operations we allow from interfaces presented to user-space are also doable from within the kernel.

Mathieu