Re: perfomance question about bool variables in cxx
From: Klaus-D. Bohn (info_at_it-bcsb.de)
Date: 07/24/05
- Next message: Keith Cayemberg: "Re: Platform Support vs. Business Support"
- Previous message: Klaus-D. Bohn: "Re: perfomance question about bool variables in cxx"
- In reply to: Bob Gezelter: "Re: perfomance question about bool variables in cxx"
- Next in thread: prep_at_prep.synonet.com: "Re: perfomance question about bool variables in cxx"
- Reply: prep_at_prep.synonet.com: "Re: perfomance question about bool variables in cxx"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Date: Sun, 24 Jul 2005 14:51:25 +0200
"Bob Gezelter" <gezelter@rlgsc.com> schrieb im Newsbeitrag
news:1122058234.888333.17740@f14g2000cwb.googlegroups.com...
> Larry,
>
> With all due respect, I disagree with your explanation. If Klaus were
> dealing with a large array of booleans, I would agree with you to the
> extent that stepping through 8 times as much data (quadwords instead of
> bytes) would generate inefficiencies.
>
> In this case, the posted code generates a large number (42,949,672 to
> be precise) of references to a single variable (quadword or byte,
> respectively). In the case of the Alpha, Itanium, and almost all VAX
> models, presuming that the data is naturally aligned by the compiler,
> each reference should be a single cache or memory reference.
>
> However, this simple experiment also illustrates the difficulties and
> challenges of benchmarking.
>
> - The compiler will, in many cases, optimize out the invariant
> assignment code in the loop. Thus, with the correct compiler switches
> used, the nearly 43 million iterations of the loop will be optimized to
> a single instruction, executed exactly once.
>
> - The fine level timing of this code, if it executes even a significant
> number of iterations will be influenced by the other activity on the
> CPU. Any event which affects cache hit rate, will effect the overall
> performance of the code. To do this experiment effectively requires a
> deeper analysis, and many test runs to get a significant certainty that
> the numbers are not being influenced by random events. Science is
> reproduceable, happenstance is not.
>
> - There is a certain amount of processing that is not attributable to
> the actual code under test. This includes image activation, the jacket
> code in the test program, and image rundown. How much this is needs to
> be computed, and subtracted from the test results to obtain an accurate
> result. This overhead is the computational equivalent of the "tare
> weight", the weight of an empty container when placed on a scale.
>
> My apologies for running on a bit. I hope that this commentary is
> helpful.
>
> - Bob Gezelter, http://www.rlgsc.com
>
Hello Bob,
thank you very much.
I have a very interesting new application for the OpenVMS platform. It will
be available in october this year.
The application is ready, but I'm looking how i could increase the
perfomance. So i make some short test programs to look what is the really
difference between the variable definitions.
Normaly i could set the counter to 1. But the system is to fast to analyze
the time difference.
I get some very interesting results. Now i chance my application to a clean
64bit application.
Next step is to analyze the application with some performance tools. I hope
i have a positive result.
Klaus
- Next message: Keith Cayemberg: "Re: Platform Support vs. Business Support"
- Previous message: Klaus-D. Bohn: "Re: perfomance question about bool variables in cxx"
- In reply to: Bob Gezelter: "Re: perfomance question about bool variables in cxx"
- Next in thread: prep_at_prep.synonet.com: "Re: perfomance question about bool variables in cxx"
- Reply: prep_at_prep.synonet.com: "Re: perfomance question about bool variables in cxx"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Relevant Pages
|
