|
|
|
         
|
Applications PerformanceResults of running a standard NCOM ocean model case on Neo, Janeway, and a few other machines. The Fortran compiler does not yet know about the SN1 (i.e. you can't use f77 -Ofast=ip35), so further optimization may be available later. The single processor time is in-line with the 300 to 400 MHz speedup. (Note that it is still significantly slower than Alpha processors.) It was very difficult to get 0.5GB of memory for a single processor job from Miser on Janeway, and there were much slower times than reported here on one processor - probably because the allocated memory was not local to the processor (or the local memory bandwidth was shared by another job). This suggests that the smaller per-processor memory capacity of Neo is going to be a significant factor on NRL's workload. Taking the 16-processor time as the most representative of scaling (since it probably isn't optimal to use all the processors of the SN1 on one job), the SN1 is about 2x as fast as the 300MHz O2K and 3x as fast as the T3E. As usual, SHMEM scales better than MPI but both MPI and SHMEM performance improves on the SN1. The following were runs on Janeway with ssusage: janeway 183> pwd /scratch/wallcraf/NCOM1/med12h janeway 184> ll *log -rw-r--r-- 1 wallcraf 28638 Jul 3 10:59 ncom1_sgimpisr04.log -rw-r--r-- 1 wallcraf 31421 Jul 3 11:01 ncom1_sgimpisr16.log -rw-r--r-- 1 wallcraf 35142 Jul 3 11:05 ncom1_sgimpisr32.log -rw-r--r-- 1 wallcraf 26991 Jul 3 10:56 ncom1_sgione.log -rw-r--r-- 1 wallcraf 28425 Jul 3 11:13 ncom1_sgishmem04.log -rw-r--r-- 1 wallcraf 30111 Jul 3 11:14 ncom1_sgishmem16.log -rw-r--r-- 1 wallcraf 32335 Jul 3 11:15 ncom1_sgishmem32.log janeway 185> cd SN1_400 /scratch/wallcraf/NCOM1/med12h/SN1_400 janeway 186> ll *log -rw-r--r-- 1 wallcraf 28536 Jul 1 08:51 ncom1_sn1mpisr04.log -rw-r--r-- 1 wallcraf 31241 Jul 1 08:52 ncom1_sn1mpisr16.log -rw-r--r-- 1 wallcraf 34740 Jul 1 08:55 ncom1_sn1mpisr32.log -rw-r--r-- 1 wallcraf 26920 Jun 30 17:54 ncom1_sn1one.log -rw-r--r-- 1 wallcraf 27916 Jul 2 12:22 ncom1_sn1shmem04.log -rw-r--r-- 1 wallcraf 29242 Jul 2 12:22 ncom1_sn1shmem16.log -rw-r--r-- 1 wallcraf 31019 Jul 2 12:57 ncom1_sn1shmem32.log Wall time for 20 time steps of NCOM (NRL Coastal Ocean Model) on a 432x208x31 domain (requires about 0.5GB of memory). A) Single processor times.
Machine types:
B) Sorted by machine type, ordered by 16-processor speed. Parallel method is either SHMEM or MPI using sendrecv. SN1 times are on a 32-processor beta machine.
Table of standard NLOM benchmark times on Neo compared to O2K's and the T3E (Note that the 300MHz O2K times quoted here were run on a SGI benchmark machine (not Janeway). The 195MHz times are from Odyssey when it was 128 processors.) This is a larger problem than the NCOM benchmark, and it is also a completely realistic run (not just timing a few time steps). The benchmark was described in the Fall NAVO MSRC Navigator, see: http://www.navo.hpc.mil/cgi-bin/Navigator/navigator.cgiTaking 28-processor times as representative, the SN1 is 1.5x as fast of the O2K-300 (in-line with clock speed) and 2x as fast as the T3E. Note that is it 2.8x faster than the 195MHz O2K, so probably the 8MB cache on the 300 and 400 MHz R12000s give a boost for this code over the 4MB cache on the 195MHz R10000. Performance of NRL Layered Ocean Model on HPC Platforms
View Halo Benchmark Graph |
Send comments or questions to ccshelp@cmf.nrl.navy.mil
NRL ~ Code 5000 ~ Code 5500 ~ Code 5590
