Skip to contentSkip to conversation

[[email protected]: Good news for glibc.]

[email protected]: Good news for glibc. Grant Grundler 3 participants 4 messages Jan 5, 2004

lore
Share & tools
lore
#0 lore raw

glibc profiling?
This was originally posted on debian-hppa/parisc-linux mailing lists.

grant

----- Forwarded message from Carlos O'Donell <[email protected]> -----

Old-Return-Path: <[email protected]>
X-Original-To: [email protected]
From: Carlos O'Donell <[email protected]>
To: John David Anglin <[email protected]>
Cc: [email protected],
[email protected], [email protected]
Subject: Good news for glibc.
X-Useless-Header: listen, love, accept, mindfullness, patience, and humility.
Resent-Message-ID: <AW1MyC.A.LG.3eQ-_@murphy>
Resent-From: [email protected]
X-Mailing-List: <[email protected]> archive/latest/3162
List-Id: <debian-hppa.lists.debian.org>
List-Post: <mailto:[email protected]>
List-Help: <mailto:[email protected]?subject=help>
List-Subscribe: <mailto:[email protected]?subject=subscribe>
List-Unsubscribe: <mailto:[email protected]?subject=unsubscribe>
List-Archive: <http://lists.debian.org/debian-hppa/>
Resent-Sender: [email protected]
Resent-Date: Mon, 5 Jan 2004 00:44:07 -0600 (CST)
X-Virus-Scanned: by amavisd-new-20030616-p5 (Debian) at lackof.org
X-Spam-Checker-Version: SpamAssassin 2.60 (1.212-2003-09-23-exp) on
colo.lackof.org
X-Spam-Status: No, hits=-4.9 required=5.0 testsºYES_00 autolearn=ham

	version=2.60

X-Spam-Level:

Quoted 4 lines · John David Anglin
Quoted 2 lines

The main issue was that the 'start' symbol needs to be canonicalized before passing on to the gmon routines. This requires entry.h to be

I still think that using __c_f_f_c for this special case is overkill.

What could we use?

Good news, I just fixed LD_PROFILE so I can actually check to see what functions take up too much time :) I think I inadvertently fixed it for ia64 too... what do ia64 people use to profile glibc?

c.

A tempting peek! :)

% cumulative self self total

 time   seconds   seconds    calls  us/call  us/call  name
  0.00      0.00     0.00    14000     0.00           cfree
  0.00      0.00     0.00     6000     0.00           memset
  0.00      0.00     0.00     2004     0.00           malloc
  0.00      0.00     0.00     2003     0.00           flockfile
  0.00      0.00     0.00     2003     0.00           funlockfile
  0.00      0.00     0.00     2001     0.00           __cxa_finalize
  0.00      0.00     0.00     1002     0.00           index
  0.00      0.00     0.00     1000     0.00           __GI__dl_close
  0.00      0.00     0.00     1000     0.00           __GI__dl_open
  0.00      0.00     0.00     1000     0.00           printf
  0.00      0.00     0.00     1000     0.00           puts
  0.00      0.00     0.00     1000     0.00           strlen
  0.00      0.00     0.00        1     0.00           __libc_start_main
  0.00      0.00     0.00        1     0.00           realloc
  0.00      0.00     0.00        1     0.00           strrchr
Signature · 4 lines
To UNSUBSCRIBE, email to [email protected]
with a subject of "unsubscribe". Trouble? Contact [email protected]

----- End forwarded message -----
#1 lore raw
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Grant Grundler

Grant> glibc profiling? This was originally posted on Grant> debian-hppa/parisc-linux mailing lists.

I'd recommend q-tools. It will give you a (statistically collected) call-graph, in addition to the flat profile and is completely unintrusive (no recompilation etc needed). See my post for details:

http://marc.theaimsgroup.com/?l=linux-ia64&m7075994721581

Oh, yeah, I'm slightly biased... ;-)

The one limitation is that it doesn't work well for short-lived processes. Also, it works in system-wide mode (collects info for _all_ processes that happen to run over a given period of time).

If you want per-process profiling, you could use qprof. It doesn't do call-graphs (yet) and only works for dynamically linked programs, but it's very easy to use. See:

http://www.hpl.hp.com/research/linux/qprof/

	--david
#2 lore raw
In reply to Grant Grundler
Quoted 7 lines · David Mosberger

I'd recommend q-tools. It will give you a (statistically collected) call-graph, in addition to the flat profile and is completely unintrusive (no recompilation etc needed). See my post for details:

http://marc.theaimsgroup.com/?l=linux-ia64&m7075994721581

Oh, yeah, I'm slightly biased... ;-)

Indeed! Looks very cool though.

Quoted 9 lines

The one limitation is that it doesn't work well for short-lived processes. Also, it works in system-wide mode (collects info for _all_ processes that happen to run over a given period of time).

If you want per-process profiling, you could use qprof. It doesn't do call-graphs (yet) and only works for dynamically linked programs, but it's very easy to use. See:

http://www.hpl.hp.com/research/linux/qprof/

Looks good, It was just my wish to get LD_PROFILE working, what's easier than tacking on an environment variable to get a quick look at the calls made by a particular DSO?

I'm 99.9% sure if upstream takes my patches this will fix LD_PROFILE for ia64 aswell.

c.

#3 lore raw
In reply to Grant Grundler
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Carlos O'Donell

Carlos> Looks good, It was just my wish to get LD_PROFILE working, Carlos> what's easier than tacking on an environment variable to get Carlos> a quick look at the calls made by a particular DSO?

Carlos> I'm 99.9% sure if upstream takes my patches this will fix Carlos> LD_PROFILE for ia64 aswell.

That would definitely be cool. The more performance tools, the merrier.

	--david