[Beowulf] AMD Opteron Linpack Benchmark

[Craig Tierney]

Lai Dragonfly wrote:
> Hi,
> There is a rank 367 which achieve 79.97% efficiency.
>  
>  
> i'm using Infiniband to connect 2 nodes. each node has 8G memory and 
> dual AMD opteron 250.
> i just tried Intel Compiler + GOTO, it just got a 60% efficiency which 
> have a large difference with my target.
> maybe i need to try Pathscale or PGI compiler.
> any suggestion?
> thanks a lot.
>  

It isn't the compiler.  Use gcc and Goto and you will get the same
result.  The biggest determiner of performance for HPL is the 
configuration of HPL.dat.  Also, run HPL on a single node to make
sure you are getting around 90% efficiency, then run across the IB
and see how things change.

Craig




> PN
> 
>  
> 2006/9/22, Craig Tierney <ctierney at hypermall.net 
> <mailto:ctierney at hypermall.net>>:
> 
>     Lai Dragonfly wrote:
>      > Dear all,
>      >
>      > I'm doing the HPL benchmark on several nodes of AMD opteron
>     platform but
>      > may expand to hundred of nodes later.
>      > i hope to get around 80% efficiency.
> 
>     What is your interconnect?  80% may be a bit high.  You should look
>     at the Top500 list  to see what the efficiencies of other systems
>     similar to yours.
> 
> 
>      > Anyone has good suggestions of combination of compiler (appreciate if
>      > includes compiler flags) + math lib on AMD platforms?
>      > thanks a lot.
>      >
> 
>     Compiler and flags don't matter.  HPL spends about 99.999% of its
>     time in BLAS.  Use GotoBLAS, it is very fast.
> 
>     What is most important is settings in the HPL.dat file.  Use google to
>     find some good settings or someone else can provide one.  I don't have
>     access to one right now.
> 
>     Craig
>     on the list may be able
> 
> 
> 
> 
>      > PN
>      >
>      >
>      >
>     ------------------------------------------------------------------------
>      >
>      > _______________________________________________
>      > Beowulf mailing list, Beowulf at beowulf.org
>     <mailto:Beowulf at beowulf.org>
>      > To change your subscription (digest mode or unsubscribe) visit
>     http://www.beowulf.org/mailman/listinfo/beowulf
> 
>