[MPICH] problems in achievement of effective speedup

[David Hart]

There is latency in going to other machines, and there is overhead in  
running MPI. These are overcome if there's enough work that can be  
parallelized; if your test jobs are small, there won't be. The trick  
of good parallel performance is to make the test job very big.

There may be other factors, but without details they cannot be known.


On Jun 21, 2007, at 11:59 AM, Anna Mereu wrote:

> Hi,
> I'm running a 5-processes application composed by one master and  
> four slaves.
> I want to run them in in parallel on two machines: the first  
> machine (machine A) has 1 processor,the second one (machine B) has  
> four processors.
> So, my intention is to run the master process in the machine A,  
> while the slaves in the machine B.
> I'm using the globusrun command with a specification file .rsl.
> The problem is that when i run the whole application only in the  
> machine A i obtain a stopping time
> smaller than if i run the master in A and the slaves in B. How can  
> it be possible?
> I know that we obtain real benefits from parallelization if we  
> maximize the parallel part of the application.
> I think that my example shoul be fine to  obtain effective speedup,  
> but in fact it doesn't work...
> Another problem is that if i run the whole application in the 4- 
> processors machine it is always much slower than in the 1- 
> processors machine.....
> So, i am very confuse about these results.....
> Can somebody help me??????
> Thank you in advance
> Anna
>