[SCore-users-jp] [SCore-users] Re: Global arrays

kameyama @ pccluster.org kameyama @ pccluster.org
2002年 6月 17日 (月) 16:38:50 JST


In article <200206141430.g5EEUJP09044 @ zeralda.streamline.com> Nick Birkett <nrcb @ streamline-computing.com> wrotes:
> I am now trying  to get SCore Global arrays working with the Portland Compile
> r
> and the Intel Compiler. This is because our academic customers who use SCore
> have these compilers.
> 
> For PGI compiler I did this (without applying your patches):
> 
> export TARGET=LINUX
> make  CC=mpicc _CC=mpicc FC='mpif77 -compiler pgi -fast' _FC='mpif77 
> -compiler pgi -fast' USE_MPI=yes LIBMPI='-lm'
> 
> (the -lm is just to reset the default from -lmpi and is not really needed).
> 
> I am using SCore 5.0.0 and 5.0.1 and Portland 3.3-2.
> 
> The above seems to work and I can run these GA  version 3-2B tests (in 
> g/global/testing)
> 
> test.x
> testsolve.x
> testeig.x
> 
> These all work.
> 
> However when I try the perf.x (performance test) it hangs at 128 byte:
> 
> 
> SCore-D 5.0.0 connected.
> <0:0> SCORE: 4 nodes (4x1) ready.
> ARMCI configured for 4 cluster nodes
>   
>  Performance of GA get, put & acc for square sections of array[1024,1024]
>   
>   
>                           Local 2-D Array Section  
>      section           get               put           accumulate
>   bytes    dim     sec      MB/s     sec      MB/s     sec      MB/s
>       8      1  .112D-05 .712D+01 .104D-05 .772D+01 .178D-05 .450D+01
>      72      3  .160D-05 .451D+02 .124D-05 .579D+02 .195D-05 .368D+02
>     128      4  .226D-05 .566D+02 .163D-05 .784D+02 .223D-05 .574D+02
> 
> Eventually I have to press <cntrl>C .

It seems to deadlock between normal process and data server.

On compute host, there are 3 processes for perf.x
parent process is called wait4() in SIGCHILD signal handler.
1 chaild process is shadow process, tis is pause() loop.
This process create by SCore, it is OK.
1 process called select().
I rewrite signal handler to remove call wait(),
following message is apears:
    0:Child process terminated prematurely, status=: 136085728
Probably, dataserver process (armci/src/dataserv.c) is dead,
but I don't know that.

                       from Kameyama Toyohisa
_______________________________________________
SCore-users mailing list
SCore-users @ pccluster.org
http://www.pccluster.org/mailman/listinfo/score-users



SCore-users-jp メーリングリストの案内