ParalellVarFiler version 1.17

[aminer]

Hello,

You will say that those scalable FIFO queues on multicores that i have presented to you are scalable in theory , this is why we call them scalable, but i don't agree, cause when you say scalable it means
scalable also in pratice, but we are far from resolving the problem
of the non scalability on multicores of the data memory transfer,
i don't mean data cache memory transfer, but data memory transfer that
is not scalable, so as i said since we can not exclude the data memory
transfer from the FIFO queue application or concept, as we can not exclude
the data memory tranfer from the parallel compression, hence
we can not say that scalable FIFO queues on multicores are scalable,
cause as i have explained to youl i think those scalable FIFO queues
are not scalable, it's the same for parallel hashtables and database
engines, we can not exclude the data memory tranfer that is not scalable
from the hashtables applications or concept or from the database engines
applications or concept, so since they are memory bound and/or disk
bound and cache unfriendly that means they are not scalable on multicores.


Thank you,
Amine Moulay Ramdane.

[aminer]

Hello,

We have to be smart please, if you take a look at my Parallel implementation of Conjugate Gradient Linear System Solver 1.02
it is here in my website:

http://pages.videotron.com/aminer/

Since the mathematic algorithm that i am using is using matrices,
so you have to  be smart and notice that you have to load the matrices
from the main thread, and give those datas inside the matrices to
other threads to do parallel calculations, but since the parallel
calculations are not so expensive as is parallel compression,
so memory's tranfers and cache misses will be more expensive
than mathematic matrices calculations, so this is why i say
that this will not scale well when the matrices become bigger,
and those kind of mathematic algorithms that uses matrices calculation don't scale well when the matrices become bigger.


Thank you,
Amine Moulay Ramdane.

[aminer]

Hello,

Question:

Amine, you have said that those scalable FIFO queues and parallel hashtables and database engines are not scalable on multicores , is
there any hope ?

Answer:

For databases engines you have to use distributed databases ,
and distribute your data over a network of computers, and
we are lucky with databases, cause databases also look like RWLocks, cause in many databases scenarios, since writes transactions are much fewer than read transactions , so the replication part that is expensive will take less time, this is why distributed databases are the answer for scalability.



Thank you,
Amine Moulay Ramdane.

[aminer]

Hello,

Question:

Amine, I think you have forgot something with distributed databases ?


Answer:

Youare right, there is still a big problem with distributed databases and that's very sad, cause distributed databases suffers from the same problem as the problem of contention in a multicore scenario, cause if there is a lot of writes transactions at the same time(as in the case of contention on multicores) if we must ensure a realtime replication
of the database, the database can suffer from high degradation in the throuhput, you will say that to resolve this problem we must use many many more computers , but this too costly i think, so that's sad.




Thank you,
Amine Moulay Ramdane.