Discussion related to cado-nfs

[Христофор Бобров <hbobrov@yandex.ru>]

Hello!
I'm using git version d1d4c2cd0f61e. In the host file are ip-address node 1 
and node 2(10.50.162.141,10.50.162.142). If I watch the top, then when I 
start on one node, processor load is 100%, and when I start at two nodes 
processor load is 50% on each. I start a job without a scheduler. What 
implementation of MPI do you recommend to use? Openmpi?


>Hi.

>Thanks for your feedback.

>This is the annoying bit:

>2 nodes, no common name prefix                                               
>   
> 10  .  .  .  .  .  .                                                        
>    
>  .  .  .  .  .  .  .                                                        
>    
> 10  .  .  .  .  .  .                                                        
>    
>  .  .  .  .  .  .  .                                                        
>    
>cpubinding: parameter ignored (no hwloc, no C++11)                           
>   

>You're essentially using the same node twice. I guess that CPU usage on
>the second node was ridiculous.

>Also, in order to enable cpu binding, you need a c++11 compiler (for gcc,
>we're happy with version 4.8 or later).

>Please make sure you use the git version, but I think it's the case.

>Note that I've never really paid attention to mpich integration. I'm
>entirely to blame about that. Hopefull your feedback from fruther trials
>will help improve this situation.

>You may want to give a try to a more recent compiler, because that should
>be the key to giving you the cpu binding code (for no good reason, really
>-- I was just lazy on that one). But that's really an aside, because the
>main oddity you're facing is elsewhere. I'm a bit puzzled by the fact
>that cado-nfs chose to put two jobs on the same node.

>What do you have in the following file ?

>/home/cado/cado_mpi/share/cado-nfs-2.2.1/dlp/host_file

>Also, is your job running from a job scheduler or not ?

>E.

29.03.2017, 20:36, "Христофор Бобров" <hbobrov@yandex.ru>:
> Hello!
> I tried to parallelize the linear phase using MPI. For the number of 87 
> decimal digits, I solved the system on one and two and three nodes, there 
> was no performance increase, for all cases the performance was "krylov: N = 
> 1000; ETA (N = 63000): [0.056 s / iter]
> I ran with cpubinding = false. During the operation, the network was not 
> 100% loaded. What could be the problem?
> I attach to the letter files with stdout bwc.pl for one and two nodes.
> --
> Best regards
> Бобров Х.

-- 
Best regards
Бобров Х.