Axel, Thank you very much for your response! > please note that benchmark numbers are useless without > offering a way to reproduce them. so please provide the > the input used for this test. > Sorry, I'm out of town right now, I'll send the input file to you directly as soon as I return home. > also, this looks almost like you are running multiple > copies of a serial binary. please check the top of the > pw.x (you _are_ pw.x, right?) output and check whether > there are the to be expected differences. > No, it's a parallel version compiled with intel MPI 3.1 and sequential MKL (libmkl_intel_lp64, libmkl_sequential and libmkl_core). The header part of my pw output indicates definitly that it's a parallel run. > finally, how does the job scale with just one pool? > The job scales quite well with just one pool, with efficiency from 92% on two cores to 74% on 8 cores, same tendency coincident with Intel's nehalem beta version.
Huiqun Zhou @Nanjing University
