Saturday, April 16, 2016

Digest for comp.programming.threads@googlegroups.com - 20 updates in 11 topics

Ramine <ramine@1.1>: Apr 15 08:32PM -0700

Hello....
 
 
He is a new idea of mine about USL..
 
When you do your testing to collect performance data
on the parallel program , if you have 4 cores,you have to test
for 1,2,3,4 cores, for example test 10 times for 2 cores, and
test 10 times for 3 cores and test 10 times for 3 cores and test 10
times for 4 cores, and calculate after that the standard deviation
between the 10 times of the tests for each case, if this standard
deviation is greater that means that you can not forecast scalability
too far with my USL programs, and that means you can forecast
scalability to about 2X times the maximum number of cores of the
performance data that is in this example 4 cores, but if the standard
deviation is small , so you can forecast scalability with my programs
too far to about 4X or 5X or even 6X times the maximum number of cores
of the performance data that is in this example 4..hope you have
understood my methodology that Dr. Gunther the author of USL didn't tell
you about.
 
So my USL programs will work efficiently this way...
 
 
 
 
Thank you,
Amine Moulay Ramdane.
bleachbot <bleachbot@httrack.com>: Apr 15 03:55PM +0200

bleachbot <bleachbot@httrack.com>: Apr 15 09:26PM +0200

bleachbot <bleachbot@httrack.com>: Apr 15 10:21PM +0200

bleachbot <bleachbot@httrack.com>: Apr 15 10:58PM +0200

bleachbot <bleachbot@httrack.com>: Apr 16 12:41AM +0200

bleachbot <bleachbot@httrack.com>: Apr 16 01:08AM +0200

bleachbot <bleachbot@httrack.com>: Apr 16 01:40AM +0200

bleachbot <bleachbot@httrack.com>: Apr 16 01:45AM +0200

bleachbot <bleachbot@httrack.com>: Apr 16 01:59AM +0200

bleachbot <bleachbot@httrack.com>: Apr 16 02:31AM +0200

Ramine <ramine@1.1>: Apr 15 08:01PM -0700

Hello,
 
 
Here is how i will explain to you USL..
 
Here is the very important thing about the USL methodology..
 
You have to know that in the USL methodology there is a
probabilistic output that is scalability, so if the parallel
part is not so greater than the serial part of the parallel program,
so there is much higher chance to score on the predicted contention and
the crosstalk, and this will show on more cores and more threads
up to 16 for example, so probabilistically , the chance is higher
to catch the right approximation of the predicted scalability
up to 3X or 4X times the maximum number of cores in the
empirical performance data, this reasonning applies to the
other cases, so a good approximation about the predicted
scalability will be at this 3X or 4X. Make the same
reasonning and you will undertand me more.
 
This is i think what we have to know about the USL methodology
that uses mathematical regression to predict scalability.
 
Thank you,
Amine Moulay Ramdane.
Ramine <ramine@1.1>: Apr 15 07:47PM -0700

Hello....
 
 
Although i have given an intuitive proof about USL in my previous post...
 
I don't think that this intuitive proof is 100% safe..
 
So the best thing is to not forecast too far with my USL programs..
 
You can forecast scalability to about 3X or 4X the number of cores
of the maximum number of cores of the empirical performance data,
and this is also useful.
 
I think that's the best thing the USL methodology can do ,
and that's the best thing my USL programs can do.
 
 
 
Thank you,
Amine Moulay Ramdane.
Ramine <ramine@1.1>: Apr 15 07:42PM -0700

Hello,
 
 
Although i have given an intuitive proof about USL in my previous post...
 
I don't thing that this intuitive proof is 100% safe..
 
So the best thing is to not forecast too far with my USL programs..
 
You can forecast to about 3X or 4X the number of cores of
of the maximum number of cores of the empirical performance data,
and this is also useful.
 
I thing that's the best thing the USL methodology can do ,
and that's the best thing my USL programs can do.
 
 
 
Thank you,
Amine Moulay Ramdane.
Ramine <ramine@1.1>: Apr 15 07:10PM -0700

Hello...
 
Dr. Gunther the author of USL didn't spookk about
what i am going to explain...
 
Please read those final thoughts that i will make more precise:
 
 
And i think that if the parallel part is variable as
in many parallel applications and is sensibly greater than
the serial part, so there is a low risk of contention and crosstalk
(represented by the Alpha and Beta coefficients of USL) at sensibly
more cores and more threads, this is a more intuitive proof, so in this
case you can forecast better the scalability with my USL programs, but
if the serial part is greater and the parallel part is not variable,
so i think the approximation of my USL programs will be useful, and if
if the serial part is greater and the parallel part is not variable
the approximation of my USL programs will be good, so since those
cases constitute almost all the cases of the parallel applications that
we encounter, so my USL programs will be good at forecasting the
scalability at more and more cores and more and more threads...
 
 
This is my final thoughts as an intuitive proof that render
my USL programs important and useful to forecast scalability.
 
So i will add to my USL programs nonlinear regression,
so stay tuned because version 3.0 of my USL programs are
coming soon...
 
 
And long and happy life to USL and my USL programs !
 
 
You can find version 2.82 of my USL programs with the source code here:
 
https://sites.google.com/site/aminer68/universal-scalability-law-for-delphi-and-freepascal
 
 
 
Thank you,
Amine Moulay Ramdane.
Ramine <ramine@1.1>: Apr 15 06:43PM -0700

Hello,
 
 
Yet another important thing about USL methodology..
 
I have thought more about USL methodology...
 
And i think that if the parallel part is variable as
in many parallel applications and is sensibly greater than
the serial part, so there is a low risk of contention at sensibly
more cores and more threads, and this to a certain point,
this is a more intuitive proof, so in this case you can forecast
better the scalability with my USL programs, but if the serial
part is greater, so the approximation of my USL programs will be good,
so since in many parallel applications that you encounter are classified
in this two cases , so my USL programs will be good at forecasting the
scalability at more and more cores and more and more threads...
 
 
So this is why my USL programs are important and useful.
 
And this why Universal Scalability Law (USL) is important and useful.
 
So i think that this approximation that i am making as a more intuitive
proof to render USL important, is useful and something important to know.
 
So this is why my USL programs are still useful, so
i will add to my USL programs nonlinear regression,
so stay tuned because version 3.0 of my USL programs are
coming soon...
 
You can find version 2.82 of my USL programs with the source code here:
 
https://sites.google.com/site/aminer68/universal-scalability-law-for-delphi-and-freepascal
 
 
 
 
Thank you,
Amine Moulay Ramdane.
Ramine <ramine@1.1>: Apr 15 05:01PM -0700

Hello........
 
Here is my final thoughts about USL ...
 
And if you look at the following web page from Dr. Gunther the
author of USL:
 
http://perfdynamics.blogspot.ca/2009/04/assessing-usl-scalability-with-mixed.html
 
He is applying his USL methodology to mixed workload on web database
systems, but i have thought more, and i think that he is wrong
to apply it this way, and i have explained why, but this is
not the end of the story, if you set constant the number
of worker threads on the database system, and apply USL
the the computer server of the database system and if
the the USL revealed a more linear scalability up the
number of worker threads, that means that we can apply
the above methodology on the link above on mixed worloads,
and i think it will work.
 
So USL methodology is still useful and important, since \
as i said before you can forecast scalability optimisticaly
with it up to 3X the number of maximum cores of the empirical
performance data as in the raytracer performance data that
i have presented, and you can use it to forecast the response-time
even on mixed workload of Ecommerce database systems and database
systems in general.
 
 
So finally i have made USL methology important to us..
 
So this is why my USL programs are still useful, so
i will add to my USL programs nonlinear regression,
so stay tuned because version 3.0 of my USL programs are
coming soon...
 
 
You find version 2.82 of my USL programs with the source code here:
 
https://sites.google.com/site/aminer68/universal-scalability-law-for-delphi-and-freepascal
 
 
 
 
Thank you,
Amine Moulay Ramdane.
Ramine <ramine@1.1>: Apr 15 04:23PM -0700

Hello,
 
I have come to an interesting subject..
 
Hope you have read my previous critics about USL..
 
I will ask now the following question:
 
Is USL methodology still useful ?
 
Here is my answer:
 
Since the USL methodology is using polynomial or
nonlinear regression on the empirical statistical
performance data, so my proof and more intuitive
affirmation [1] that i stated before says:
 
"I have just thought about his methodology of using Parallel
systems as a black box using mathematical regression to quantify
and predict scalability, and i think he is wrong doing it so,
because when the serial part is for example 1/8 of the parallel part,
you can probabilistically get much more scalability with less parallel
threads and less cores than when testing with more and more cores,
because with more and more cores you will get more and more contention
on the serial part , so the Dr. Gunther methodology fails , because
it is testing the system with less cores and less threads, so it will
not get and predict the right or a better approximation of the
scalability" [1]
 
 
But this is not the end of the story, in fact this kind
of statistical regression can for example give an optimistic forecast on
approximatly 2X or 3X of the maximum number of cores of the empirical
test that we have obtained with raytracer performance data, because
since it is a regression so it can foerecast optimistically this far.
 
So this is why my USL programs are still useful, so
i will add to my USL programs nonlinear regression,
so stay tuned because version 3.0 of my USL programs are
coming soon...
 
 
 
Thank you,
Amine Moulay Ramdane.
Ramine <ramine@1.1>: Apr 15 03:29PM -0700

Hello..
 
I explain more my critics about the USL methodology...
 
If you have read my previous post, i have based my proof on
this affimation that is more ituitive, read it carefully:
 
"I have just thought about his methodology of using Parallel
systems as a black box using mathematical regression to quantify
and predict scalability, and i think he is wrong doing it so,
because when the serial part is for example 1/8 of the parallel part,
you can probabilistically get much more scalability with less parallel
threads and less cores than when testing with more and more cores,
because with more and more cores you will get more and more contention
on the serial part , so the Dr. Gunther methodology fails , because
it is testing the system with less cores and less threads, so it will
not get and predict the right or a better approximation of the
scalability" [1]
 
 
And if you look at the following web page from Dr. Gunther:
 
http://perfdynamics.blogspot.ca/2009/04/assessing-usl-scalability-with-mixed.html
 
He is applying the USL methodology to a mixed workload to
predict the response-time etc.
 
But i think he is wrong, because there is an unknown that
is the number of cores and the number of server threads on
the server side of the database system, so you can not more
exactly predict scalability because as i said in [1] above,
you can not predict in general how much contention there is
on the serial part represented by coefficient Alpha and you
can not predict how much cross-talk there is represented by
the Beta coefficient.. so i hope that my more intuitive
affirmation [1] above is clear.
 
So i say that Dr. Gunther methodology of USL fails and is
not the right tool, and more than that , as i said before it is not
suited for critical systems that demand and require us to know
the worst-case performance.
 
Thank you,
Amine Moulay Ramdane.
Ramine <ramine@1.1>: Apr 15 09:56AM -0700

On 4/14/2016 9:09 PM, Ramine wrote:
> and predict scalability, and i think he is wrong doing it so..
 
> Because when the serial part is for example 1/8 of the parallel part,
> you can probalistically get much more scalability with less parallel
 
I mean: probabilistically, not probalistically.
 
 
You received this digest because you're subscribed to updates for this group. You can change your settings on the group membership page.
To unsubscribe from this group and stop receiving emails from it send an email to comp.programming.threads+unsubscribe@googlegroups.com.
Posted by at

No comments:

Post a Comment

Subscribe to: Post Comments (Atom)