Michael Galloway <mgx@ornl.gov> writes:

out of curiosity, how can you measure the performance of individual cpus?

I ran the LINPACK benchmark in two xterm windows concurrently (and had a third xterm window display the output from "top"). If I run one instance of linpack then every other time (on average) it will run slowly. If I start two instances of linpack at the same time, then one will run at full speed and the other at 1/10:th the speed. It's a pity that Linux "top" doesn't print which CPU a certain program is running on like it does in Solaris. I also miss the Solaris "psradm" program so I could have programmatically shut down one of the CPUs at the time in order to try to pinpoint which of the CPUs it is that is causing the problem... XTERM WINDOW #1 ----------------------------------------------------------------------- [1] opteria:~/src/linpack> out/opteria/linpack.gcc Rolled Double Precision Linpack Size of REAL: 8 bytes Array sizes: aa: 7812 Kbytes, a: 7820 Kbytes norm. resid resid machep x[0]-1 x[n-1]-1 9.5 4.22017976e-12 2.22044605e-16 1.09912079e-13 5.08926234e-13 ttimes are reported for matrices of order 1000 dgefa dgesl total kflops unit ratio ttimes for array with leading dimension of 1001 35.33 0.10 35.43 18873 0.11 632.68 31.39 0.00 31.39 21302 0.09 560.54 XTERM WINDOW #2 ---------------------------------------------------------------------- [0] opteria:~/src/linpack> out/opteria/linpack.gcc Rolled Double Precision Linpack Size of REAL: 8 bytes Array sizes: aa: 7812 Kbytes, a: 7820 Kbytes norm. resid resid machep x[0]-1 x[n-1]-1 9.5 4.22017976e-12 2.22044605e-16 1.09912079e-13 5.08926234e-13 ttimes are reported for matrices of order 1000 dgefa dgesl total kflops unit ratio ttimes for array with leading dimension of 1001 2.81 0.00 2.81 237960 0.01 50.18 2.81 0.01 2.82 237116 0.01 50.36 -- Peter Eriksson <peter@ifm.liu.se> Phone: +46 13 28 2786 Computer Systems Manager/BOFH Cell/GSM: +46 705 18 2786 Physics Department, Linköping University Room: Building F, F203 SE-581 83 Linköping, Sweden http://www.ifm.liu.se/~peter