MathGroup Archive 2000

[Date Index] [Thread Index] [Author Index]

Search the Archive

About Parallel Computing

  • To: mathgroup at smc.vnet.net
  • Subject: [mg25868] About Parallel Computing
  • From: "nikaidoh" <nikaidoh at mech.fukui-u.ac.jp>
  • Date: Sat, 4 Nov 2000 02:04:13 -0500 (EST)
  • References: <8tdrao$5uc@smc.vnet.net> <200011010625.BAA19248@smc.vnet.net>
  • Sender: owner-wri-mathgroup at wolfram.com

Jens,
Thank you very much.
Your advice was very much useful.

I doubt a command to evaluate now.
Parallel Computing will be almost successful.

> Computation time is determined by the *algorithm* and it's
> parallelisation.
> In a second look it depends on the speed of the CPU connection
> (pipe speed on SMP machines, TCP/IP speed in ws-clusters)
> 
> The Parallel Computing toolkit support only the "master - slave"
> topologie. You can still design a peer-to-peer algorithm but with out
> an additional programming your task is distributed onto all slaves and
> the master is waiting.
> 
> Since you have only two machines you can do the following to get a
> speed up. 
> a) let the faster machine act as slave
> b) start a second kernel as second slave on the slower machine *and*
>    start a kernel process as master on the slower machine
> 
> Hopefully the better conection speed on the same machine will a bit
> out weight the speed of the faster machine. Since the master is
> waiting and consume only minor CPU time the two slaves are working and
> may speed up your calculation.
> 
> 
> How ever the speed up depends on your true task and the algorithm you
> use.
> So I can't give you an advice what's to do in you case.
> 
> Hope that helps
>   Jens
> 
> 



  • Prev by Date: Re: Is it possible to plot a point a time on the same graph in Mathematica?
  • Next by Date: Important, interesting, intermediate level exercise.
  • Previous by thread: Re: About Parallel Computing
  • Next by thread: Re: combinatoric card problem