Poll on Intel VTune(TM) Performance Analyzer for Windows

[Steven_L_Intel1]

On this board there is a poll asking about use of Intel VTune Performance Analyzer for Windows. If you're an Intel Visual Fortran user, please take the poll. If you would be interested in a whitepaper or a webinar specifically oriented towards Fortran users of VTune, please reply here. Any other comments you might have about why you do or do not use VTune can be placed here as well. Thanks.

[durisinm]

I bought VTune but have used it only a few times. The learning curve is a little steep for me. I would attend a Webinar on VTune for Fortran users if one were presented. A few examples would probably go a long way toward making me more comfortable and familiar with it.

I feel that VTune's cost is on the high side, but it's probably worth it to developers producing software products for sale.

Mike D.

[keefer]

Hi,

I agree with most of what durismn said. The learning curve for effective usage is very, very steep. One not only has to learn something about the P4 assembly language, but also a significant amount about how the L2 cache and instruction pipeline works. The latter is very confusing (64K aliasing, write backs, etc.). This is in addition, of course, to learning how V-Tune itself works. Any further documentation would help. I would also recommend to Intel that they offer a stripped down version for programmers who do not program for networks etc. (Call them V-tune Lite, or V-tune Std. and V-tune Pro). V-tune is expensive. If one simply wants his code to run faster, the choice is between V-tune and a processor upgrade. At street price, upgrading to a P4 3.2Ghz Extreme Edition is only about $100 more than V-Tune's list price.

I first tried a trial version of V-tune about 2 years ago. The trial license was good for only 10 days, about 1/10th the time to get anywhere up the learning

curve if one does not already know assembly language and have a good idea how the L2 cache works. Fortunately, Intel was kind enough to offer a beta version last fall, to anyone with Pre. Sup. The license was good for about 10 weeks and I took advantage of it, although I was not capable of distinguishing bugs (if any), from user errors.

The good news is that V-tune works well, especially for IVF8.X. After much effort I discovered that 30%-40% of the CPU time was spent on an innocuous looking store instruction. I rearranged the source code so that a computationally intensive statement was put several lines ahead of the store so the cache intensiveoperation would (hopefully) occur concurrently with the CPU intensive one. It worked and the program ran almost twice as fast. Of course the source code was no longer logically arranged.

A really nice solution to the whole issue would be a profiler based on V-Tune, so that the INTEL coding experts could reduce these problems. After all, a Fotran programmer does not expect a stupid store instruction to kill his program's performance. There is absolutely no way to even gain a suspecision (sp?) of the problem from the source code itself.

I appologize for my usual verbosity.

Regards,

Keith

[Steven_L_Intel1]

Thanks for the comments, Mike and Keith. Longer is better!

Keith, have you tried the Profile-Guided Optimization feature in IVF? And did you compile with /opt:3?