Board index » delphi » Vote on IntToStr becnhmarking

Vote on IntToStr becnhmarking


2005-09-07 03:43:57 PM
delphi221
Hi all,
I would like to ask you to vote on the topic of IntToStr benchmarking.
Currently three benchmark methods are available:
(a) Current benchmark
Benchmark six subranges with uniform distribution on each
subrange. stability through inner loop / outer loop construct
such
as often found in Dennis' benchmarks. This means: measurement
over
inner loop, minimum over outer loop.
Included in current B&V release found on attachments (unchanged
from earlier releases) as default benchmark.
(b) Alternative AZ benchmark
Benchmark entire range at once with distribution proportional to
inverse of logarithm. stability through inner loop / outer loop
construct.
Included in current B&V release found on attachments; can be
enabled in defines.inc.
(c) Alternative SZ benchmark
Benchmark entire range at once with uniform distribution.
stability by repeating measurement and taking minimum. This means:
measurement over each execution, several executions with equal
parameters in a row, minimum taken over executions in a row,
sum/average taken over all minimums of executions.
Included in "Alternatiuve IntToStr speed test" found on
attachments.
I vote (b), because i think this new benchmark:
- has a realistic probability distribution
- doesn't make it easy on the branch predictor by mixing numbers of all
orders of maginitude
- performs more accurate measurement by not measuring every execution,
but the total inner loop at once (see my reply to Sasa's thread).
- is consistent with previous challenges by using the inner/outer loop
construct
Please keep the score, including names. Current vote results:
(a)
(b) AZ
(c)
If you prefer an altered version of any of these, please register your
vote with the one you want altered and make a comment about the changes
you would like.
 
 

Re:Vote on IntToStr becnhmarking

Hi
Keep the votes coming.
(a)
(b) AZ, Dennis
(c)
Regards
Dennis
 

Re:Vote on IntToStr becnhmarking

Avatar Zondertau writes:
Quote
Hi all,

I would like to ask you to vote on the topic of IntToStr benchmarking.
Currently three benchmark methods are available:
I think it is too soon for vote. My own benchmark is just early beta.
Sasa
--
www.szutils.net
 

Re:Vote on IntToStr becnhmarking

Avatar Zondertau writes:
Quote
(c) Alternative SZ benchmark

Benchmark entire range at once with uniform distribution.
stability by repeating measurement and taking minimum. This means:
measurement over each execution, several executions with equal
parameters in a row, minimum taken over executions in a row,
sum/average taken over all minimums of executions.
Not quite. That is only one posibility to for measurements. Setting one
loop per turn and it will comply branch prediction request and return real
result of executing (as other benchmarks does).
This also give stability as much with RDTSC is possible. If we would have
fair vote, please wait couple of days I finish it as I planed (have to
many regular obligations at work right now) or simply delete it from vote
list. As I mentioned already, it is still early beta.
Sasa
--
www.szutils.net
 

Re:Vote on IntToStr becnhmarking

Hi Sasa
Quote
I think it is too soon for vote. My own benchmark is just early beta.
OK. Keep the releases coming and we will hold back voting for a while.
Regards
Dennis
 

Re:Vote on IntToStr becnhmarking

Quote
>(c) Alternative SZ benchmark
>
>Benchmark entire range at once with uniform distribution.
>stability by repeating measurement and taking minimum. This
>means: measurement over each execution, several executions
>with equal parameters in a row, minimum taken over executions
>in a row, sum/average taken over all minimums of executions.

Not quite. That is only one posibility to for measurements. Setting
one loop per turn and it will comply branch prediction request and
return real result of executing (as other benchmarks does).

This also give stability as much with RDTSC is possible. If we would
have fair vote, please wait couple of days I finish it as I planed
(have to many regular obligations at work right now) or simply delete
it from vote list. As I mentioned already, it is still early beta.
Deleting it from the vote list would mean not using it, since it's
preferable to not repeat voting. However it seems fair we will wait
until you finish your benchmark suggestion.
Note BTW that it is quite easy to implement your benchmark inside the
B&V program. You can do this in the class for the challenge, like i did
with my alternative benchmark in the last release. This would make it
possible to test the benchmarks under the same circumstances and would
minimize work needed to put your benchmark suggestion in the B&V if
that would be the outcome of the vote.
 

Re:Vote on IntToStr becnhmarking

Avatar Zondertau writes:
Quote
Note BTW that it is quite easy to implement your benchmark inside the
B&V program. You can do this in the class for the challenge, like i
did with my alternative benchmark in the last release. This would
make it possible to test the benchmarks under the same circumstances
and would minimize work needed to put your benchmark suggestion in
the B&V if that would be the outcome of the vote.
I'm a bit short with spare time at the moment to examine the B&V
source. Also, I do not prefer 6 grups as they did and by my opinion
need to be changed to test every digit group. That require examination
of the source for which unfortunatelly I do not have a time. I will
continue work on my current release for now.
Sasa
--
www.szutils.net