[mpich-discuss] MPI how to support the smp machine?

[liaoyin]

If  I add the N , the peak performance do not go up.

I use the netpipe(http://www.scl.ameslab.gov/netpipe/) and IMB-MPI to test
the communication between two core in one cpu.

with no diskless platform.

I run mpirun -np 2 ./NPmpi  and mpirun -np 2 ./IMB-mpi in a node (two
prcesses run on the two core), the bandwidth is high and latency is small

but with diskless.

I run mpirun -np 2 ./NPmpi and mpirun -np 2 ./IMB-mpi in a node (two
prcesses run on the two core), the bandwidth is very low and latency is
large.


2009/12/6 Gustavo Correa <gus at ldeo.columbia.edu>

> Hi Liaoyin
>
> Your problem size N=1000 is too small.
> The order of P,Q also matters.
> See my previous message for suggestions.
>
> Gus Correa
>
>
> On Dec 5, 2009, at 8:06 AM, liaoyin wrote:
>
>  thank you .
>> my mpi is mpich2 1.2.
>>
>> My cluster is a  diskless platform that is load kernel and  mount
>> filesystem from the server.
>> Single node  has  one cpu with four core.(the theoretical peak performance
>> in single core is 3.2 Gflops )
>>
>> I am runing the  HPL ( A Portable Implementation of the High-Performance
>> Linpack Benchmark for Distributed-Memory Computers) to do linpcke test.
>>
>> I use the Atlas to as a blas lib that is a single thread version.
>>
>> case 1, I run the mpirun -np 1 ./xhlp   the Gflops is 0.4 ( I found that
>> one core to run with the top command  HPL.dat is that n=1000,ps*qs=1*1)
>>
>> case 2, I run the mpirun -np 2 ./xhlp   the Gflops is 0.04  (I found that
>> one core to run with the top command HPL.dat is that n=1000,ps*qs=1*2)
>>
>> I use the netpipe(http://www.scl.ameslab.gov/netpipe/) and IMB-MPI to
>> test the communication between two core in one cpu.
>>
>>
>> with no diskless platform.
>>
>> I run mpirun -np 2 ./NPmpi  and mpirun -np 2 ./IMB-mpi in a node (two
>> prcesses run on the two core), the bandwidth is high and latency is small
>>
>> but with diskless.
>>
>> I run mpirun -np 2 ./NPmpi and mpirun -np 2 ./IMB-mpi in a node (two
>> prcesses run on the two core), the bandwidth is very low and latency is
>> large.
>>
>> why?
>>
>>
>> 2009/12/5 Gus Correa <gus at ldeo.columbia.edu>
>> Hi Liaoyin
>>
>> Besides Dave Goodell's questions,
>> I wonder if this may be more of an HPL problem
>> than of an MPICH2 problem.
>>
>> Here are some questions/suggestions about HPL:
>>
>> ***
>>
>> 1) If you just want to check if MPICH2 is working,
>> then don't use HPL.  Use something simpler.
>> The cpi.c and hellow.c programs in the MPICH2 "examples" directory
>> will tell you if MPICH2 is working properly, and are way much
>> simpler than HPL to setup and run.
>>
>> ***
>>
>> 2) Note, there may be a typo on the mpirun command line
>> on your message, it should be "xhpl" not "xhlp" as you wrote,
>> unless you changed the executable name.
>>
>> ***
>>
>> 3) Are you sure you are running HPL on 2 cores?
>>
>> If you use Linux, you can submit your HPL job,
>> then use "top" (and type "1") to see how many
>> cores are actually running xhpl.
>>
>> Do you really see 2 cores in action on "top"?
>>
>> ***
>>
>> 4) What are the contents of your HPL.dat parameter file,
>> when you try to run on 2 cores?
>>
>> Is is the same that you use for the one core run or is it different?
>>
>> This may not be news to you, but here it goes just in case:
>>
>> To run HPL on a given number of cores on your machine,
>> the product of the values or Ps and Qs in HPL.dat has to be
>> equal to the N value on your "mpirun -np N"
>> For example, if you want to use 2 cores (mpirun -np 2),
>> you could use this on your HPL.dat file:
>>
>> 1 2 Ps
>> 2 1 Qs
>>
>> Or to use four cores:
>>
>> 1 2 4 Ps
>> 4 2 1 Qs
>>
>> (Tip: You *must* use a *single blank space* field separator in HPL.dat.)
>>
>> ***
>>
>> 5) Very small problem sizes don't scale well with
>> the number of processors.
>>
>> To see speedups when more cores are added,
>> you need to choose a relatively large value for Ns in HPL.dat.
>> However, Ns is constrained by how much memory (RAM) your computer has.
>> A reasonable upper bound to Ns is sqrt(0.8*memory_in_bytes/8).
>>
>> See item 5) below.
>>
>> ***
>>
>> 6) Finally, if you haven't read it, the HPL TUNING file is a
>> *must read* for anybody who wants to run the HPL benchmark:
>>
>> http://www.netlib.org/benchmark/hpl/tuning.html
>>
>> ***
>>
>> Good luck!
>>
>> I hope this helps,
>> Gus Correa
>> ---------------------------------------------------------------------
>> Gustavo Correa
>> Lamont-Doherty Earth Observatory - Columbia University
>> Palisades, NY, 10964-8000 - USA
>> ---------------------------------------------------------------------
>>
>>
>>
>>
>> Dave Goodell wrote:
>> What MPI implementation are you using?  If you are using MPICH2, please
>> make sure that you are using the latest version (1.2.1).
>>
>> -Dave
>>
>> On Dec 4, 2009, at 3:02 AM, liaoyin wrote:
>>
>> I am runing the  HPL ( A Portable Implementation of the High-Performance
>> Linpack Benchmark for Distributed-Memory Computers) to do linpcke test.
>>
>> I use the Atlas to as blas lib.
>>
>> My machine has  one cpu with four core.
>>
>> I run the mpirun -np 1 ./xhlp   the Gflops is 0.4  (one core to run)
>>
>> but  I run the mpirun -np 2  ./xhlp the Gflops is 0.04 (two core to run)
>>
>> why is the two core is slower.
>>
>> _______________________________________________
>> mpich-discuss mailing list
>> mpich-discuss at mcs.anl.gov
>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>>
>> _______________________________________________
>> mpich-discuss mailing list
>> mpich-discuss at mcs.anl.gov
>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>>
>> _______________________________________________
>> mpich-discuss mailing list
>> mpich-discuss at mcs.anl.gov
>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>>
>> _______________________________________________
>> mpich-discuss mailing list
>> mpich-discuss at mcs.anl.gov
>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>>
>
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20091206/d38ab365/attachment-0001.htm>