Reuti <[email protected]> writes: >> This site http://dsavas.staff.shef.ac.uk/teaching/sge.pdf talks about using: >> >> -l arch=intel* >> -l arch=amd* > > I assume they overwrote simply the default setting in the exechost > definition (`qconf -me node01` => complex_values arch=foobar). But I > wouldn't suggest to do so,
Indeed. > but to introduce a new complex "mtype" or whatever and fill it with an > appropriate value (it could even be put in a load sensor to set it > automatically and there might be more possibilities, whether the > machines have sse4 or not and for fma3/4). I have a complex "cputype" with (6!) values like "interlagos" and "westmere". If you're doing machine-specific builds, cache and scheduling parameters are probably relevant for performance, even if they don't stop the binary running. > Some packages like Gaussian provide different binaries for each CPU > type and depending on your jobscript you could either request a > preferred one or chose the right one automatically during execution > (if you put your binaries in congruent directories where you select > the type of CPU at an upper level). I'd do the latter, especially as you potentially want to dynamically link different BLAS etc., as in the previous post. -- Community Grid Engine: http://arc.liv.ac.uk/SGE/ _______________________________________________ users mailing list [email protected] https://gridengine.org/mailman/listinfo/users
