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

Groups > comp.programming > #3872 > unrolled thread

About Lockfree...

Back to article view | Back to comp.programming

Contents

  About Lockfree... aminer <aminer@toto.net> - 2013-10-07 20:07 -0700
    Re: About Lockfree... aminer <aminer@toto.net> - 2013-10-07 20:14 -0700
    Re: About Lockfree... aminer <aminer@toto.net> - 2013-10-07 20:35 -0700

#3872 — About Lockfree...

Hello,

We have to be smart please, there are those that
like Lockfree algorithms, but in my opinion those
lockfree algorithms have an important weakness, in high contention
scenarios they  will higher the cache-coherence traffic
so they will slow the things, in worst scenarios you can
have 1+2+3+4+ ..N = ((N^2+N)/2) or even worse N+N+N+N...N = N^2
cache lines transfers from the memory and this is not good at all,
so what i will advice is to wrap those push() and pop() of your FIFO 
queue with a Lock that supports FIFO fairness and that lower the 
cache-coherence traffic, this is better i think, so you can use for example
the Scalable Anderson Lock , the scalable MCS Lock or my scalable 
distributed Lock for that , those Locks are FIFO fair and  better for 
realtime systems.



Thank you,
Amine Moulay Ramdane.

[toc] | [next] | [standalone]

#3873

I wrote:
 > so what i will advice is to wrap those push() and pop() of your FIFO
 > queue with a Lock that supports FIFO fairness and that lower the
 > cache-coherence traffic, this is better i think, so you can use for 
example
 > the Scalable Anderson Lock , the scalable MCS Lock or my scalable
 > distributed Lock for that , those Locks are FIFO fair and  better for
 > realtime systems.


And it's better to use two Locks , one for the push() and one for the pop().


Thank you,
Amine Moulay Ramdane.




On 10/7/2013 8:07 PM, aminer wrote:
>
> Hello,
>
> We have to be smart please, there are those that
> like Lockfree algorithms, but in my opinion those
> lockfree algorithms have an important weakness, in high contention
> scenarios they  will higher the cache-coherence traffic
> so they will slow the things, in worst scenarios you can
> have 1+2+3+4+ ..N = ((N^2+N)/2) or even worse N+N+N+N...N = N^2
> cache lines transfers from the memory and this is not good at all,
> so what i will advice is to wrap those push() and pop() of your FIFO
> queue with a Lock that supports FIFO fairness and that lower the
> cache-coherence traffic, this is better i think, so you can use for example
> the Scalable Anderson Lock , the scalable MCS Lock or my scalable
> distributed Lock for that , those Locks are FIFO fair and  better for
> realtime systems.
>
>
>
> Thank you,
> Amine Moulay Ramdane.
>
>

[toc] | [prev] | [next] | [standalone]

#3874

On 10/7/2013 8:07 PM, aminer wrote:
> We have to be smart please, there are those that
> like Lockfree algorithms, but in my opinion those
> lockfree algorithms have an important weakness, in high contention
> scenarios they  will higher the cache-coherence traffic
> so they will slow the things, in worst scenarios you can
> have 1+2+3+4+ ..N = ((N^2+N)/2) or even worse N+N+N+N...N = N^2
> cache lines transfers from the memory


Those lockfree FIFO queue algorthms higher the cache-coherency traffic 
when there is contention, in worst scenarios you can
have 1+2+3+4+ ..N = ((N^2+N)/2) or even worse N+N+N+N...N = N^2
cache lines transfers from the memory and those Lockfree FIFO queue
algotithms are not FIFO fair , hence you can also have starvation with 
Lockfree FIFO queue algorithms, and this is not good at all,
so what i will advice is to wrap those push() and pop() of your FIFO 
queue with two Locks that supports FIFO fairness and that lower the 
cache-coherence traffic, this is better i think, so you can use for 
example the Scalable Anderson Lock , the scalable MCS Lock or my 
scalable distributed Lock for that , those Locks lower the 
cache-coherence traffic and are FIFO fair , so they avoid starvation and
are better for realtime systems.




Amine Moulay Ramdane.

[toc] | [prev] | [standalone]

Back to top | Article view | comp.programming

csiph-web