dtrace discuss - Jun 2007 - Instrumentation of programs using DTrace USDTs

Had a chat with a fellow Sun Engineer on this. Here we do not completely agree.
Wrote a blog and am wondering how others would think about this. Of course there
must be other opinions here.. Does someone already have experience with a full
blown monitoring attempt of a non trivial application using USDTs?

If you would like to read the blog:
http://blogs.sun.com/paulvandenbogaard/entry/monitoring_why_when_what

Would really value your contribution on this topic.
--
This message posted from opensolaris.org

Hi Paul,

I have just a couple of comments on your post. First, you mention context --
which is a bit of a coy way of identifying that DTrace is a _systemic_
tracing facility rather than one that''s application specific. Not only
does
it give you some specific context, but it allows a user to instrument any
component of the system.

Second, I think you''ve identified an important point in that probes
should
have the potential to lead to actionable conclusions short of modifying the
code. Developers can already use the pid provider (or fbt in the kernel) so
there''s no need to target them. It''s important to build
knowledge on how to
instrument the system and interpret the results in order to intelligently
turn the various knobs the application exposes.

Decicing on the right probes for an application, library or subsystem can
be quite tricky. This is why we encourage any developer who''s working
on
USDT instrumentation to post to dtrace-discuss and ask for comments from
the DTrace community.

On a related subject, Brendan Gregg and I have been discussing the idea of
''generic'' providers or ''application''
provider so that applications in a
specific domain could implement and conform to a given provider in addition
to providing their own probes. For example, Oracle, Postgres and MySQL
should all have their own provider which expose some of their own nuances,
but then they could all implement a common ''database'' provider
with some
least-common-denominator probes. This would allow the same scripts to be
used across all databases. We''ll send mail to the list when we have
something
a bit more concrete.

Adam

On Mon, Jun 25, 2007 at 11:05:53AM -0700, Paul van den Bogaard
wrote:
> Had a chat with a fellow Sun Engineer on this. Here we do not completely
agree. Wrote a blog and am wondering how others would think about this. Of
course there must be other opinions here.. Does someone already have experience
with a full blown monitoring attempt of a non trivial application using USDTs?
> 
> If you would like to read the blog:
> http://blogs.sun.com/paulvandenbogaard/entry/monitoring_why_when_what
> 
> Would really value your contribution on this topic.
> --
> This message posted from opensolaris.org
> _______________________________________________
> dtrace-discuss mailing list
> dtrace-discuss at opensolaris.org
-- 
Adam Leventhal, Solaris Kernel Development       http://blogs.sun.com/ahl