[Beowulf] Opteron, ifort and kernels
Many of your questions may have already been answered in earlier discussions or in the FAQ. The search results page will indicate current discussions as well as past list serves, articles, and papers.
Timothy W. Moore twm at tcg-hsv.comWed Aug 10 15:04:22 PDT 2005
- Previous message: [Beowulf] MPI_Open_port: difficult questions (I think)
- Next message: [Beowulf] 90nm opteron power usage
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Hi Patrick - I have 3 Opteron clusters and an Itanium 2 cluster. I have already performed (or at least tried) such benchmarks with the exception of Intel-generated code running on the opteron or 64 bit AMD code running on I2. The Itanium 2 chip is a RISC (no x86) chip (AMD64 is capable of X86 instruction sets) and codes optimized for I2 are unable to execute outside of an I2 platform. What I am driving at is that AMD64 will not run Intel 64 bit (and vice versa)...different optimizations for different chipsets. Your best bet is to take your best shot at the I2 with the Intel compilers. Next compile the code in 32 bit mode and run it on the opteron. And finally, get the PGI or Pathscale compilers and take your best shot at AMD64 and compare the results. I have found that on AMD64 and 32 bit processors that Intel generated code is about 10-15% faster than PGI. Pathscale has not worked for me due to some directives in FORTRAN being unspoorted. Do not use GCC on the I2...one of our CFD applications runs at less than half the speed of the same code compiled with the Intel V7 compilers. That comment leads to yet another about the compilers...I sometimes wonder if Intel understands how to develop compilers that will compile optimized code for the I2. We have several programs that will compile but not run, or if they do run, are very slow. Version 7 of the compilers were great in that they generated some beautiful, highly-optimized CFD code (only one though). Version 8 took the same code and made garbage of it. I am currently looking at 9 to see if it is any better. Intel states that V9 is more compatible with V7 and preliminary results look good. My AMD64 clusters run mixtures of gcc/PGI-generated code in 32 and 64 bits. Unfortunately, I have future requirements that will force me to run 32 bit Intel-generated code on AMD64. I have an Intel server box networked to the AMD64 rack for just such purposes. This is my long answer to a short question!!! I hope my knowledge and experience with the various compiler families has helped...I find it rare to be of assistance ... the topics are sometimes out in left field. If someone has experience different from that of my own, I would appreciate and welcome any contradictory information. Tim Chief Scientist, TCG On Wed, 2005-08-10 at 12:00 -0700, beowulf-request at beowulf.org wrote: > Message: 1 > Date: Tue, 09 Aug 2005 09:52:27 +0200 > From: Patrick Begou <Patrick.Begou at hmg.inpg.fr> > Subject: [Beowulf] Opteron, ifort and kernels > To: beowulf at beowulf.org > Message-ID: <42F860BB.5040907 at hmg.inpg.fr> > Content-Type: text/plain; charset=us-ascii; format=flowed > > Some weeks ago I saw an interresting discussion about opteron anf intel > fortran. > > Please could you let me know about your configurations ? > > - linux distribution and kernel version ? > - ifort version (32 or 64 bits ?) > > A friend provide me a login on an opteron based server for runing some > tests but when I want to install Intel fortran compiler (8.1.) it is the > 32 bit version which get installed! > > I would like to make some comparison between my itanium based Altix and > an opteron based server with same compiler version. > > Thanks for your advices > > Patrick
- Previous message: [Beowulf] MPI_Open_port: difficult questions (I think)
- Next message: [Beowulf] 90nm opteron power usage
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the Beowulf mailing list
