[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-smp
Subject:    Re: possible use of a linux cluster (fwd)
From:       Manoj Warrier <manoj () plasma ! ernet ! in>
Date:       2000-02-29 6:56:37
[Download RAW message or body]

Hi.

> I use this occasion to propose something to the linux-smp world : when you
> want to send a MPI message within a SMP node, almost all of the MPI
> implementation use a shared memory area and do 2 copies (Memory space
> Process 1 => Shared memory Area => Memory space Process 2). This design is
> OK for small messages, the latency is good if you avoid the use of locks.
> But for larege messages, the bandwith is limited to half of the memory
> copy performance (2 copies) and the shared memory area used is large.
> I have implemented a direct copy with linux using a kernel module to copy
> the data from User space 1 to user space 2 in one copy. The bandwith
> measured is 2 times the initial bandwith.
> 
> It would be great that the linux kernel itself provides this function. Do
> you think it's possible/impossible/usefull/useless ?
> 
> Patrick GEOFFRAY
> ---
> Matra Systemes & Information - MS&I
> Ecole Normale Superieure de Lyon - LIP
> Universite Lyon I
> http://lhpca.univ-lyon1.fr

I was looking forward for replies from all u ppl qualified enough to
comment on the above proposal. Me being a newbie to kernel hacking
and wanting to explore the possibility of running a linux cluster with
dual processors, I feel it is useful, but I really cannot comment on
the possibility/impossibility/desireablity of including this in the
Linux kernel.

-Manoj

e-mail : manoj@plasma.ernet.in
fax    : 0091-79-2864310

post   : Manoj Warrier, Scientist SC,
	 Institute for Plasma Research,
	 BHAT, Gandhinagar, Gujrat,
	 India - 382428

-
Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/dmentre/smp-howto/
To Unsubscribe: send "unsubscribe linux-smp" to majordomo@vger.rutgers.edu

[prev in list] [next in list] [prev in thread] [next in thread]