[MPICH] Running IMB-IO with MPICH2

[Jothi Padmanabhan]

Hi Rob,

I am seeing the problem with NP=2, Aggregate itself.  Here is a
cut-paste of the error I am seeing...

Regards
Jothi

#------------------------------------------------------------------------------------------
# Benchmarking P_Write_Indv
# #processes = 2
#------------------------------------------------------------------------------------------
#
#    MODE: AGGREGATE
#
       #bytes #repetitions  t_min[usec]  t_max[usec]  t_avg[usec]
Mbytes/sec      defects
            0           50        52.92        52.92        52.92
   0.00         0.00
            1           50        50.78        57.38        54.08
   0.02         0.00
            2           50         4.00         7.76         5.88
   0.25         0.00
            4           50         3.54         7.90         5.72
   0.48         0.00
            8           50       769.80      1141.46       955.63
   0.01         0.00
           16           50       852.20       939.48       895.84
   0.02         0.00
           32           50       829.74       844.70       837.22
   0.04         0.00
0: Error Write_xxx, size = 64, sample #38
Error: restored buffer from output file, invalid portion starting at pos. 2464
Process 0: Erroneous data:
Buffer, 1st and last entry: 0 0


On 6/25/07, Robert Latham <robl at mcs.anl.gov> wrote:
> On Mon, Jun 25, 2007 at 03:33:32PM +0530, Jothi Padmanabhan wrote:
> > I am trying to run Intel MPI Benchmarks (IMB_3.0) with mpich2-1.0.5p4.
> > I am getting a data integrity error (I am running with -DCHECK option).
> Has
> > anybody faced a similar problem?
> >
> > Here are the details:
> >
> > IMB-IO, P_Write_Indv benchmark,
> > NP=2 on a single node,
> > UFS,
> > Fedora Core 4,
> > 32-bit Pentium 4 processor.
>
> Hi Jothi
>
> Was this aggregate or non-aggregate mode?
>
> I don't see any problems with either Aggregate or non-aggregate mode
> P_Write_Indv at 2 processors.   At 4 processors, I also have no
> problems with aggregate mode.
>
> 4 processor non-aggregate P_Write_Indv, however, crashes really early
> in the benchmark.   That's pretty odd to me, because it's my
> understanding (section 5.2.8.2 of the IMB manual) that the only
> difference between aggregate and non-aggregate mode is that buffers
> are flushed after every operation in the non-aggregate case.
>
> ==rob
>
> --
> Rob Latham
> Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
> Argonne National Lab, IL USA                 B29D F333 664A 4280 315B
>