symlinking ld to the correct one solves that problem:
Code:
# ll /usr/bin/ld* |
Quote:
Facundo |
Hello,
currently, I'm testing OF on two Intel Xeon PHI cards using a native OF build but I cannot reproduce the results of the benchmark obtained by asaijo. I get really bad results for PHI architecture. Did anyone know what code modifications should be done to get the performance improvements described in this thread? Best regards Matthias |
Hi
I followed the setup here as shown from kokizzu but still lots of error messages as : ~OpenFOAM/OpenFOAM-2.2.2/wmake/platforms/linux64Icc/wmkdep: cannot execute binary file .. can any one tell me how to fix that? |
Hello:
I also met this problem, but here my "compilervars_arch.sh" don't have option "mic", I am very sure we have several xeon Phi cards on each node here. is it because of installation of compiler? thanks! Quote:
|
Hi everybody, thank you so much for sharing your Phi/OpenFoam sucesses and difficulties.
I'd like to try the same for naval hydrodynamics and have a 8*Xeon Phi machine at hand (Supermicro 4028, Dual Haswell, eight 31S1P mic's, accessible via Teamviewer). I want the solution to run on the Phi's only, cluster-in-a-box-style: The host OS and processors would only care about the virtual networking and storage then. No need to care about the complicated offloading/coprocessing stuff. I'm both an OpenFOAM and C++ nOOb, and already overloaded with other parts of the project, so it's probably hopeless for me to get this setup to work all by myself. So i'm looking for someone with significant experience who would be interested in a small consulting mission. First step would be to compile OpenFOAM for the mic's, maybe perform some tuning along what others have tried (hints here: https://www.nersc.gov/assets/Uploads...5-OpenFOAM.pdf) and then benchmark a reference case (~20 million cells) against the same load running on the two Haswell host processors. I guess this is only a few days of work for someone who knows his stuff, and it would allow to decide if it's worth pursuing in terms of raw machine performance. (I'd be equally interrested in that kind of test for SU2 if you are familiar with that.) If that succeeds (success meaning that one phi card can more or less do the work of a Haswell CPU), the second step would be to engage in workflow setup. Between the discussions about requirements, installation, testing, writing up (nothing formal needed, i just need to memorize what to do and what to keep in mind) and explaining i'd expect this to take around a couple weeks of work. A typical marine case would be 2DOF or 6DOF with or without waves, including some kind of turbulence modelling. Apart from the large hull 15m there will also be small apendages (think .5 to 3m) behaving like airfoils. Results will be graphical pressures, velocities, flowlines on the hull, and tables with resistance, position, accelerations. I don't necessarily need absolute data because i can calibrate against known stuff. Context: I'm a lonesome entrepreneur doing the feasibility homework of my startup project. At this stage i have more or less proven my design (on a commercial cfd package) and am now into optimizing it. Later on i'll have it cross-checked by more knowledgable/better equipped hydrodynamics consultants. So there is no big money involved at this stage, and even when actually launched my project would become a small business with a handful of people. However the months of sheer cost of living and computing expenses start to pile up to significant amounts, and since you could save me months i'd be happy to honor that. We could of course also talk about machine time if you have heavy calculations to do yourself. Also i'll probably have other stuff to do later on if we get on well. The thread is probably not the best place to discuss details so please private message me or drop me a mail at davidtimide <at> gmail <dot> com if interested! Thanks for your kind attention and a nice day to all! |
Hi All,
When i build openfoam for xeonphi, I am getting the following error. Can some one please tell me what could be the problem. ld: Relocatable linking with relocations from format elf64-k1om (Make/linux64IccDPOpt/sigFpe.o) to format elf64-x86-64 (/share_nfs/openfoam-xeonphi/OpenFOAM/OpenFOAM-1.7.x/lib/linux64IccDPOpt/libOSspecific.o) is not supported make: *** [/share_nfs/openfoam-xeonphi/OpenFOAM/OpenFOAM-1.7.x/lib/linux64IccDPOpt/libOSspecific.o] Error 1 + wmake libso OpenFOAM wmakeLnInclude: linking include files to ./lnInclude /bin/sh: /share_nfs/openfoam-xeonphi/OpenFOAM/OpenFOAM-1.7.x/wmake/bin/linux64Icc/wmkdep: cannot execute binary file /bin/sh: /share_nfs/openfoam-xeonphi/OpenFOAM/OpenFOAM-1.7.x/wmake/bin/linux64Icc/wmkdep: cannot execute binary file /bin/sh: /share_nfs/openfoam-xeonphi/OpenFOAM/OpenFOAM-1.7.x/wmake/bin/linux64Icc/wmkdep: cannot execute binary file /bin/sh: /share_nfs/openfoam-xeonphi/OpenFOAM/OpenFOAM-1.7.x/wmake/bin/linux64Icc/wmkdep: cannot execute binary file |
Hi All,
The above linking problem resolved after adding LD flags(-melf_k1om) in make file. Thanks Sivaramakrishna |
Quote:
I appear to be having the same problem (error linking from format elf64-k1om to elf64-x86-64). Are you able to elaborate on how exactly you added that flag? For example, what file it's added to and the syntax? Thanks, Tom |
Hi,
Add -melf_k1om flag to LD_FLAGS in make files c, c++. If that does not work change symbolic link of ld to ld of mic architecture to resolve linking errors. Thanks Sivaramakrishna |
Quote:
Code:
wmake/rules/linux64Icc Doesn't work for me. Neither does it work when I say in Code:
etc/bashrc Code:
export WM_CXXFLAGS=$WM_CFLAGS Code:
./Allclean |
Compiling easy? Not at all!
Hi Guys,
Quote:
Question: Do I need to compile this freaky thirdParty Metis, ParMGridGen, Libccmio and stuff in order to make OpenFoam or Foam-Extend usable? For me, I solved many issues for the compiling procedure (Cross-compilation for Xeon Phi Co-Processor 3110) resulting in a super long manual, but with this Metis and libccmio Stuff I run into strange linker issues that seem to stem from the insufficient configure scripts for qmake that do not implement my -mmic option but use their own with -m64. When I try to compile Foam-Extend without this third-party stuff, libfoam.so compiles incomplete and so does the rest of course. I would like to leave all this thirdparty stuff out because fixing the rpm_make routine --- ahem --- s*cks. Or do I do a mistake? --- Edit: Ok, I left out the ThirdParty stuff and I was able to compile at least the sources /src without scotch, metis and parmidgren. But going for the applications, I also get this error: Code:
k1om-mpss-linux-ld: warning: libimf.so, needed by /home/max/foam/foam-extend-4.0/lib/linux64IccDPOpt/libinterfaceProperties.so, not found (try using -rpath or -rpath-link) Code:
#echo $LD_LIBRARY_PATH |
All times are GMT -4. The time now is 19:51. |