From mboxrd@z Thu Jan  1 00:00:00 1970
From: compudj@krystal.dyndns.org (Mathieu Desnoyers)
Date: Thu, 22 Jan 2009 23:49:40 -0500
Subject: [ltt-dev] lttng development plan
In-Reply-To: <20090123213907.34DC.KOSAKI.MOTOHIRO@jp.fujitsu.com>
References: <4973F6CD.5040008@cn.fujitsu.com> <20090120015553.GD9224@Krystal>
	<20090123213907.34DC.KOSAKI.MOTOHIRO@jp.fujitsu.com>
Message-ID: <20090123044940.GB24971@Krystal>

* KOSAKI Motohiro (kosaki.motohiro at jp.fujitsu.com) wrote:
> > > >   - Integrate LTTng with LKCD, test with kernel crash extraction,
> > > >     create tools to simplify extraction of traces from crashed kernel,
> > > >     integrate those tools to ltt-control.
> > > 
> > > We will implement it.
> > > 
> > 
> > Great :) Note that there has already been some work done on this. This
> > in available as an addition to the crosscrash tool :
> > 
> > http://sourceforge.net/projects/crosscrash/
> > 
> > There is a cross-crash-ltt.patch file available on the project website,
> > but I think it has not been updated since 2007. Some integration work
> > will have to be done.
> > 
> > Also outputting the traces in the video card's memory would be a
> > nice-to-have, because this memory often survives hot reboots.
> 
> Just dumb question.
> 
> Do you receive this feature requesting from anybody?
> 
> Currently, expected flight-recored mode user is enterprise guys.
> They usually use crash dumps. then, kernel panic cause tons i/o anyway 
> although tracing infomation don't.
> 
> Then, this feature doesn't reduce rebooting times, I think.
> 

Yes, people in enterprise need to be able to diagnose _why_ the system
failed. Failing once is acceptable from a customer perspective, but
failing again isn't. In this case, being able to extract the last events
before the crash can be very valuable and helps solving the problems
before they happen again.

I agree with you that short reboot times are also very important in
production, and we need to find a solution which ideally could provide
both :
- short reboot time
- availability of tracing data across reboot

I think that using memory such as what is available in video cards
(which survives hot reboots) could be a solution to this : the system
could reboot as soon as it crashes, and the trace would be available to
be extracted from the machine once the system is back online (given that
we put the next trace buffers at different memory addresses of the video
card).

Long reboot times seems to be a core problem with the crash dump
approach. Do you think the video card memory scheme could become
attractive enough that people would need it ?

Mathieu

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68