Cluster Performance
On May 15th, 2006, the University of Florida High Performance Computing Center submitted to the Top 500 Super Computer Sites a performance score of 2.975 TeraFLOPs, using 206 nodes with 832 CPUs. This consisted of 204 machines with four processors each (816 CPUs) and two machines with eight processors apiece (16 CPUs).
The HPC Center has not done more tests since the initial HPL test on the Phase IIa cluster in 2006. These tests take a lot of time and our users usually prefer to use the machine rather than boast about its performance in some benchmark. There is no point in testing Phase III by itself and we will not get the whole machine to test, so we may not get a new test before Phase III comes online either.
This test scales pretty well, though. The 2006 number is for 800 CPUs, so the current 1,600 CPU machine is about 6 Tflops and with 896 CPUs added, Phase III will be about 9 Tflops.
This does not put us in any position in the top 500 list that is worth mentioning, so the HPC Center feels it is not justified doing the test.
The result was achieved using a version of HPL compiled with the GNU compiler suite, the opteron-specific GOTO library, and OpenMPI. We ran one MPI task per CPU (four per node).
The output of the HPL run can be viewed here.
