VTune analyzer 2.0 beta test archives

[jeffrey-gallagher]

This readonly thread is an archive of the postings that were posted during the beta test program for this software. The product is now currently available for sale or evaluation at:

http://www.intel.com/software/products/vtune/vlin

cheers

jdg

[walles]

Using VTL2.0 rc2, build 235:

"
$ vtl view a1::r1 -code -mn libjvm.so -fn mmParScanObject
VTune Performance Analyzer 2.0 for Linux*
Copyright (C) 2000-2003 Intel Corporation. All rights reserved.

Wed Sep 24 15:56:17 2003 ERROR: critical error when initalizing source view data provider:Unexpected problem has occured.
An error occurred during navigation.
"

This error message isn't very helpful. What does it mean, and what can I do about it?

Cheers //Johan

[jeffrey-gallagher]

Great question Johan. I haven't seen it before myself, I'll pass it over to the engineers and see what they say.

Stay tuned.

cheers

jdg

[walles]

-hotspot-function doesn't work for listing samples per kernel function:

"
$ vtl view -hotspot-function -mn vmlinux
VTune Performance Analyzer 2.0 for Linux*
Copyright (C) 2000-2003 Intel Corporation. All rights reserved.

Event Summary
Clockticks
120830 = Samples collected due to this event
1779000 = Sample after value used during collection
214956570000 = Total events (samples*SAV)

HotSpot View grouped by function (all values in decimal except where indicated)

No symbols were found
"

Am I doing something wrong, or is this a known problem?

OProfile (http://oprofile.sf.net) manages to do this with the vmlinux file as input, so it should be doable.

Cheers //Johan

[jeffrey-gallagher]

GREAT question, Johan. I'm asking the relevant engineers.

cheers

jdg

[jeffrey-gallagher]

Hey Johan,

Here's the reply from engineering:

the error is probably due to not specifying full path to the le. the error message was replaced with an error message saying the module can not be located.

If that fixes the problem, let us know back here.

IF NOT, you should:

1) pack up the project that shows the problem (will create a file ending in .vxp)

2) go to premier.intel.com, log in, and open a case

3) attach the .vxp file to the case

Our engineering folks eat this stuff for breakfast...

:)

cheers

jdg

[jeffrey-gallagher]

Hi Gang,

Many thanks to all who joined the VTune Analyzer 2.0 for Linux* beta test program! The program is officially over, and the released version of the software hit the press at LinuxWorld in NYC last week, January 21, 2004, and is available for download or ordering now.

If you're still operating beta software under a beta license, it will kick out on JAN 30, so you should upgrade to the released product as soon as possible, here:

http://www.intel.com/software/products/vtune/vlin/

(By 'upgrade' I mean either buy or download an evaluation copy.)

Just a quick reminder: the VTune Analyzer 2.0 for Linux* software has the following new features:

-- callgraph for Itanium processor Linux*
-- source view
-- pause/resume APIs
-- more supported OS versions
-- a new getting started guide


Remember, the evaluation version of the software is fully functional, although that license will time out in 7 days from the day you recieve the license.

ALSO, if you are a fulltime student or an instructor who may beinterested in our related pricing plans (Academic Pricing, Student Pricing, or Classroom Pricing) you can check them outat the URL below. Each licensing scheme is slightly different, but you know which category applies to you when you read their descriptions:

http://www.intel.com/software/products/global/academic.htm

Happy tuning!

cheers,

jdg

Message Edited by jdgallag on 01-28-2004 11:09 AM

Message Edited by jdgallag on 01-28-2004 11:09 AM

[walles]

Using the full path to libjvm.so helped. Thanks!

[walles]

I've done a profiling run using the event type "Branches Retired". Here's an assembly view (slightly censored to fit) of the (start of the) result:

"
$ vtl view -code -mn /localhome/johan/clients/johan/ariane/build/linux32-j2se1.4.2/release/libjvm.so -fn mmParScanObject -asm|head -20
VTune Performance Analyzer 2.0 for Linux*
Copyright (C) 2000-2003 Intel Corporation. All rights reserved.

Legend
Ev1 = Branches Retired samples
Ev1 ICACHE Flag Decoder... Source
1 0x72 subl $0x2c, %esp
0 0x72 pushl %ebp
0 0x72 pushl %edi
1 0x72 pushl %esi
0 0x72 0x6a pushl %ebx
1 0x72 0x6e movl 0x40(%esp), %ebp
1 0x72 0x70 movl (%ebp), %eax
45 0x72 movl 0x4c(%eax), %edx
1 0x72 0x6a movl %edx, 0x28(%esp)
1 0x72 0x6e testb $0x1, %dl
0 0x72 0x70 je B_aad17
0 0x72 0x6a addl $-4, %esp
0 0x72 0x70 movl 0x48(%esp), %edx
0 0x71 0x6a pushl %edx
"

Now, how come a movl instruction gets hit by the "Branches Retired" event? I thought that one should trigger on branch instructions only.

Cheers //Johan

[mwittle]

I've had luck with
vtl view -hotspot-function -mn vmlinux -sd /usr/src/linux
after recompiling the kernel with "-g".

It seems like all I should have needed was a vmlinux that was not stripped, but I didn't check that before recompiling.

The quick start guide gives info on howto recompile with "-g", but if your vmlinux file has symbols (nm /usr/src/vmlinux), try the above command line. If it works, please let me know. If it doesn,t try recompiling the kernel with "-g".

[mwittle]

Is Vtune 2.0 for Linux supported on Red Hat Enteprise Linux 3.0 beta 2?

[jeffrey-gallagher]

Not yet.

Intel policy is not to support beta releases, until they're fully cooked and released.

cheers

jdg

[jeffrey-gallagher]

GOOD job, Walles, and thanks for reporting back.

cheers

jdg

[jeffrey-gallagher]

Excellent question, DB!

Absolutely JAVA support is in there now. Read about all the new features:

WHAT'S NEW

Support for several new Linux* distributions, including their default versions of kernel, glibc, libstdc++, and glib, has been added to VTune analyzer 2.0 for Linux*. The new Itanium processor distribution supports Red Hat* Linux* WS 2.1 and ES 2.1. Also, a newly supported IA-32 Linux* distribution has been added, SuSE* Linux* 8.2.

Java* application profiling on supported IA-32 Linux* distributions has been added. Running Java* applications with the VTune analyzer with either sampling or call graph profiling options enabled obtains results similar to native VTune profiling. Java* Virtual Machines (JVMs) from BEA*, IBM*, and Sun* are supported. Note that call graph profiling tracks Java* calls only, and there is currently no support for mixed Java* and native calls in this release. Additionally, there is currently no support for profiling Java* applications on Itanium processor versions of Linux*.

NOTE: See the Release Notes (included with the beta software) for a full list of supported Linux* distributions and Linux* JVMs.

Several differences between this software and the VTune analyzer v. 1.1 are also worth noting:

-- VTune analyzer now provides call graph output on the flow of control of Itanium processor Linux* applications in addition to event-based sampling.

-- The new Getting Started tutorial teaches the basic concepts and functionality of the VTune analyzer for Linux*.

-- Source View is now available.

-- VTune analyzer 2.0 can now import other sampling data files (*.tb5) for viewing.

[jeffrey-gallagher]

DBricker asked a great question, and I want everybody to see it.

[Daniel_B_Intel2]

Jeff,

Does VTune Analyzer 2.0 for Linux* supports data collection for Java applications?

Thank you
Daniel

[jeffrey-gallagher]

Walles,

This is a great question. If you find that you are thirsty for more information on this topic in general, don't hesitate to examine the Reference Manual:

/opt/intel/vtune/doc/reference/index.htm

BUT, to answer your question here and now, the thing to remember is, "branches retired" is not a "precise" event. This means there will be some event skid after the event occurs, and before the VTune interrupt stops the processor (which is just a part of the way the VTune sampling methodology works).

In your example, it looks like the event skid is about 8-10 instructions -- the top of the code looks like the start of a procedure (setting up the stack frame?) -- so the top instruction or the one before it is probably the target of a branch that is causing the event.

ALSO, let's take a step back here a second and think about what branching does: it jumps. The skid on branches is accordingly very unpredictable. It could even be MORE than 8-10, it could be a HUGE number, depending how far you are jumping.

In a really bad case, you might not get the vtune interrupt until the processor executes the instruction AFTER the branch. By then the IP will be where the code branched to, which could be very, very far from where you were.

Take a look at the reference guide: I bet you'll enjoy it.

cheers

jdg

[walles]

That worked. Thanks!

[jeffrey-gallagher]

Excellent.

[jayaprakashp]

Hi,

Just downloaded the beta 2.0. I was going through the VTLRELEASENOTES.htm. It mentioned that "Do NOT install this software on a Linux* system that currently has the VTune Performance Analyzer Linux* Remote Agent on it". Unfortunately, that is what I wanted to do...

I want to install the 2.0, and see if that helps me collect traces better than the remote agent (with 7.0 on a windows m/c). I dont want to mess too much with my current setup as I use the remote agent to get all data.

Can you suggest a way by which I can safely test this on the same machine and see if it will of use? Or atleast, can you tell me what happens if I unload the remote agent driver module and then do the 2.0 installation?

After installation, I want to make sure I can still use remote agent if I want to collect some serious data/or if I think I cannot use 2.0 for a particular case.

Thanks,
JP

[jeffrey-gallagher]

Hey jayaprakashp,

Do not install VTune analyzer 2.0 for Linux beta on any server with a functioning VTServer, VTune 7.0 on it (ditto for 7.1 beta, as well). There is currently no supported solution for what you want to do.

You make a great case for us changing our current decision on this; however, that certainly won't be happening in time for this beta program.

Note that the installation of VTune analyzer for Linux (vtl, either 1.1 or the 2.0 beta) will definitely corrupt the installation of the VTune analyzer (VTServer, either 7.0 or the 7.1 beta). Neither would then work, you'll have bits and pieces of both and not one happy camper between them.

To correct, from there you'd end up having to uninstall both, and then reinstall the one you wanted to use (in your case, the 7.0 Linux Remote Agent).

Depending on how you chose to uninstall, your project files could still be around, but it's easy to pick the wrong choice in the heat of an uninstall. I'd urge you strongly to avoid this experiment completely.

But, I am hearing what you're saying, that you'd like to be able to do it, and I'm forwarding that information on.

I won't explain our decision here except to say at current prices -- $699 USD a pop -- the VTune team never envisioned that users would buy both versions of the software and install on the same server.

The case you're describing, kicking the beta tires, is certainly one exception. Another might be in the classroom, with limited hardware, and where both software versions might be taught.

Keep these great comments coming! But for now, you'll have to find another server to kick the tires.

cheers

jdg