Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.programming > #3889 > unrolled thread

Scalability...

Back to article view | Back to comp.programming

Contents

  Scalability... aminer <aminer@toto.net> - 2013-10-09 18:17 -0700
    Re: Scalability... Robert Wessel <robertwessel2@yahoo.com> - 2013-10-09 20:42 -0500

#3889 — Scalability...

Hello,

So if you have undersood my ideas and my arguments aginst the Amdahl's 
equation, i have proved to you that the Amdahl's equation gives you the 
scalability that you will have in an IDEAL CONTENTION SCENARIO , but if 
you have less contention it will scale much better, and in a none 
contention scenario it will have a perfect scalability, so if you have a 
small serial part
let say of S = 0.1% so the Amdahl's equation says that your program will
scale to a maximum of 10, but that's false , it can scale much more
than that if you introduce randomness in the Parallel part to avoid
the ideal contention scenario, something like 
wait(random(something_like_the+time_of_your_bigger_locked_region)) or 
something like that and this will scale your application to much more 
than 10X, so the very important part is to introduce an optimal 
randomness of a more optimal wait , but you don't have to use sleep() 
for that,
just to loop around a pause asm instruction for example and this will give
you better scalability than 10X.


Thank you,
Amine Moulay Ramdane.

[toc] | [next] | [standalone]

#3891

On Wed, 09 Oct 2013 18:17:43 -0700, aminer <aminer@toto.net> wrote:

>
>Hello,
>
>So if you have undersood my ideas and my arguments aginst the Amdahl's 
>equation, i have proved to you that the Amdahl's equation gives you the 
>scalability that you will have in an IDEAL CONTENTION SCENARIO , but if 
>you have less contention it will scale much better, and in a none 
>contention scenario it will have a perfect scalability, so if you have a 
>small serial part
>let say of S = 0.1% so the Amdahl's equation says that your program will
>scale to a maximum of 10, but that's false , it can scale much more
>than that if you introduce randomness in the Parallel part to avoid
>the ideal contention scenario, something like 
>wait(random(something_like_the+time_of_your_bigger_locked_region)) or 
>something like that and this will scale your application to much more 
>than 10X, so the very important part is to introduce an optimal 
>randomness of a more optimal wait , but you don't have to use sleep() 
>for that,
>just to loop around a pause asm instruction for example and this will give
>you better scalability than 10X.


Your argument is silly.  If 10% (and let's use the percent sign
correctly - .1% is .001, or one part in a thousand, 10% is .1, or one
part in 10) of the work has to be done sequentially, then you cannot
hope to finish in less than 10% of the time it takes to run the
workload sequentially.  So the maximum theoretical speedup is the
inverse of that, or 10 times (1/.1).

And I'm not sure what you think an "IDEAL CONTENTION SCENARIO" is, or
why it needs to be shouted, but Amdahl's law is assuming that there is
zero contention in the parallel work (contention is essentially
modeled as serial work).  Zero contention is pretty ideal, in my book,
and unlikely to be reducible.

[toc] | [prev] | [standalone]

Back to top | Article view | comp.programming

csiph-web