This is a test to obtain the general Numpy performance.
This test goes many different python2 examples and then reports an average time. These subtests run of varying length. Following is example output
reverse_cumsum Python 2410 2466 38 diffusion Python 21340 21383 43 mandel Python 297265 302523 4935 local_maxima Python 31333 31447 123 fft Python 9103 9297 100 fdtd Python 725573 731770 4118 rosen Python 10088 10182 54 evolve Python 1631 1710 38 cronbach Python 829 845 10 grayscott Python 37203 37440 142 wave Python 27278 27832 749 vibr_energy Python 1479 1490 17 julia Python 771131 786176 16515 growcut Python 705226 714857 11488 periodic_dist Python 872 877 7 repeating Python 949 961 11 multiple_sum Python 1598 1609 18 wdist Python 28934 29183 145 hyantes Python 107016 112128 5632 allpairs_distances Python 14829 15087 601 arc_distance Python 971 976 4 harris Python 3718 3760 29 smoothing Python 325344 327078 1047 l2norm Python 1806 1836 27 hasting Python 3 3 0 allpairs_distances_loops Python 16262 16326 77 create_grid Python 3383 3408 37 pairwise Python 1310533 1312147 1503 slowparts Python 2187 2194 7 grouping Python 1621 1683 33 conv Python 1008397 1008954 540 lstsqr Python 5261 5279 18 specialconvolve Python 4242 4269 20
What gets printed is the sum of the average times (last column), so a mix of some long-running processes with very short ones.
The test is single-threaded and all testing was done bound to core 1.
Metrics (Intel) - phoronix/numpysh - pid 17444 On_CPU 0.125 On_Core 1.000 IPC 2.430 Retire 0.500 (50.0%) FrontEnd 0.277 (27.7%) Spec 0.154 (15.4%) Backend 0.069 (6.9%) Elapsed 2566.96 Procs 39 Maxrss 52K Minflt 7526820 Majflt 0 Inblock 0 Oublock 648 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 274 (19.7%) Nivcsw 1114 Utime 2561.043313 Stime 5.910712 Start 142662.17 Finish 145229.13
The code is On_Core 100%. There are some frontend stalls, but overall a medium IPC.
Process Tree - phoronix/numpy
Process Tree
The process tree is simple as it goes through each workload.
17444) sh
17445) numpy
17446) run.sh
17447) dirname
17448) python2
17454) python2
17455) python2
17456) python2
17457) python2
17458) python2
17459) python2
17468) python2
17469) python2
17470) python2
17471) python2
17472) python2
17473) python2
17474) python2
17475) python2
17476) python2
17477) python2
17478) python2
17480) python2
17481) python2
17482) python2
17486) python2
17487) python2
17524) python2
17525) python2
17526) python2
17527) python2
17533) python2
17534) python2
17548) python2
17549) python2
17551) python2
17554) python2
17555) cat
17556) awk
Metrics (AMD) - phoronix/numpysh - pid 5298 On_CPU 0.062 On_Core 1.000 IPC 1.897 FrontCyc 0.093 (9.3%) BackCyc 0.268 (26.8%) Elapsed 3423.38 Procs 39 Maxrss 51K Minflt 7543159 Majflt 0 Inblock 0 Oublock 664 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 274 (0.1%) Nivcsw 320371 Utime 3417.064884 Stime 5.136417 Start 19908.09 Finish 23331.47
AMD IPC just a bit lower.
Scheduled 100% of the time.
Slight variations in IPC amongst workloads.
Frontend stalls are the largest issue.
on_cpu 0.125 elapsed 2490.642 utime 2481.306 stime 2481.306 nvcsw 967 (36.59%) nivcsw 1676 (63.41%) inblock 0 inblock 1240 retire 0.645 ms_uops 0.005 speculation 0.031 branch_misses 38.82% machine_clears 61.18% frontend 0.261 idq_uops_delivered_0 0.036 icache_stall 0.003 itlb_misses 0.001 idq_uops_delivered_1 0.093 idq_uops_delivered_2 0.158 idq_uops_delivered_3 0.235 dsb_ops 50.94% backend 0.063 resource_stalls.sb 0.005 stalls_ldm_pending 0.520 l2_refs 0.006 l2_misses 0.001 l2_miss_ratio 18.50% l3_refs 0.000 l3_misses 0.000 l3_miss_ratio 8.84%
Frontend limited, though not necessarily icache or itlb (branch resteers or inefficient uop generation?).
Next steps: Understand why AMD IPC is lower.