Xiaodan Tang(deleted)
|
RE: TDP, SMP, and Grid computing
|
Xiaodan Tang(deleted)
10/14/2008 1:13 PM
post14972
|
RE: TDP, SMP, and Grid computing
We can do "remote spawn()" but what you are asking, is basically a
"remote fork()". :)
In fork() case, not only the fds but all the memories need to be all
copied. Based on how big the application is,
this (sending X MB of memory) may just defeat the purpose.
I think it still need a "framework". Have a "distributor" dispatch jobs
into different clusters. And have worker/distributor
talk to each other to decided who's busy/who's not...
The "APS" on this networked environment is basically controlled by the
"distributor" (where to dispatch the job).
-xtang
> -----Original Message-----
> From: Mario Mastrodicasa [mailto:community-noreply@qnx.com]
> Sent: Tuesday, October 14, 2008 12:47 PM
> To: ostech-core_os
> Subject: Re: TDP, SMP, and Grid computing
>
> It's right that some work must be done in order to achieve
> these objectives, but this kind of OSes is desirable in other
> environment (not necessary in embedded world) where cluster
> of computer are used.
> Note that some UNIX Oses have framework to achieve this objective.
> As is written in the QNX site for an AMP architecture:
>
> * Run different instance of QNX Neutrino on each cores
> * Preserve legacy operation
> * QNX Transparent Distributed Multiprocessing for
> communication between cores
>
> It is possible that it do that until now, but if is possible
> extends Transparent Distributed Multiprocessing to multiple
> CPU using the same approach that TDP apply to the resources
> (all the elements visible on the filesystem) of each node I
> think it is possible to create an AMP system using multiple
> remote CPU.
> The APS, that works on multi-core CPU, can do the same job.
> In an SMP system /proc folder is shared between all cores (or CPU).
> In an AMP I think that each OSes is joined to one cores and
> have its own /proc folder. Well, in the last scenario all
> OSes can see the /proc of the other nodes (/net/<node
> name>/proc). "Moving" /proc subfolders, and adapting opened
> file descriptor (in order to point to the correct resource),
> to a less loaded CPU all the work is done.
>
> Mario.
>
> _______________________________________________
> OSTech
> http://community.qnx.com/sf/go/post14971
>
>
|
|
|