Alan's Unified Model benchmarks page

Unless otherwise specified, the platform is a Beowulf cluster with dual-processor 1GHz Pentium 3 nodes, and Myrinet interconnect, and code is run at 32-bit using the Intel Fortran Compiler with -O3 optimisations.

This page under development. List items which are not hyperlinks are still to come:


HadAM3 and HadCM3L scaling

Config. HadAM3 Stash off HadCM3L Stash off HadCM3L Stash on
Mins / model day Speedup c.f. 1x1 Model years / tot CPU days Mins / model day Speedup c.f. 1x1 Model years / tot CPU days Mins / model day Speedup c.f. 1x1 Model years / tot CPU days
1x1 6.45 1.00 0.62 6.68 1.00 0.60 6.34 1.00 0.63
2x1 3.20 2.01 0.62 3.69 1.81 0.54 3.76 1.69 0.53
1x2 3.19 2.02 0.63 3.66 1.83 0.55 3.74 1.69 0.53
2x2 1.69 3.81 0.59 1.90 3.52 0.53 1.97 3.22 0.51
2x3     1.33 5.01 0.50    
4x2 0.91 7.09 0.55 1.03 6.47 0.48 1.06 5.96 0.47
2x4 0.89 7.21 0.56 1.02 6.54 0.49 1.06 6.01 0.47
1x8     1.06 6.31 0.47    
3x3 0.81 7.91 0.55 0.94 7.09 0.47 0.98 6.44 0.45
4x3     0.75 8.87 0.44    
2x6     0.75 8.95 0.45    
4x4 0.5012.88 0.50 0.6011.10 0.42 0.6310.06 0.40
4x5 0.4414.62 0.45 0.5412.34 0.37    
4x6 0.3916.61 0.43 0.4913.64 0.34    
5x5     0.5113.02 0.31    
3x9     0.4614.51 0.32    
2x14     0.4614.48 0.31    
8x4     0.5113.18 0.25    
4x8 0.3319.29 0.37 0.4415.23 0.28 0.4514.14 0.28

graph
of model throughput

Notes


Compiler comparison

The following are for HadAM3, 2x2 MPP, no stash. The Portland compiler is up there with the Intel Compiler if you include appropriate performance flags. (NB as stated above, this is for Pentium 3 chips. Obviously the Intel compiler might be less optimised with non-Intel CPUs.)

CompilerOptimisationsmins / model day
Intelifc -tpp6 -O3 -unroll1.65
ifc -tpp6 -O2 -unroll1.69
Portland Grouppgf90 -O2 -Munroll -Mnoframe -Mvect=sse1.67
pgf90 -O21.92
Lahey-Fujitsulf95 -O --tpp1.90

Also some single-processor HadAM3 tests (with much stash output - see standard HadAM3 job in UMUI).

CompilerOptimisationsmins / model day
Intelifc -O27.57
Portland Grouppgf90 -O28.49
Lahey-Fujitsulf95 -O8.83