This is a test of GraphicsMagick with its OpenMP implementation that performs various imaging tests to stress the system’s CPU.
The Graphics Magick code has five operations:
GraphicsMagick 1.3.28:
pts/graphics-magick-1.7.0
Processor Test Configuration
1: HWB Color Space
2: Blur
3: Local Adaptive Thresholding
4: Resizing
5: Sharpen
These are run in the following order: blur, sharpen, resize, color space and adaptive thresholding. The first four run with processes on each core, the adapting thresholding has only a single copy.
Metrics (Intel) - phoronix/graphics-magickphoronix-test-s - pid 31760 On_CPU 0.300 On_Core 2.399 IPC 1.479 Retire 0.820 (82.0%) FrontEnd 0.153 (15.3%) Spec 0.136 (13.6%) Backend -0.108 (-10.8%) Elapsed 949.47 Procs 392 Maxrss 231K Minflt 40568424 Majflt 0 Inblock 0 Oublock 1072 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 27022 (54.4%) Nivcsw 22623 Utime 2229.506677 Stime 47.962057 Start 583656.30 Finish 584605.77
First some observations about metrics for all the workloads together: (1) the On_CPU is only 30% and this has 50% voluntary context switches. While there isn’t a large amount of I/O, this does suggest that processes are mostly stopped and not scheduled on the cores (2) the topdown metrics assume something running on all cores and hence reason they look bogus above; more useful to look at workload by workload (3) the overall IPC above is slightly lower but consistent with other workloads. Looking at these as a single-threaded program “fixes” the topdown metrics overall:
On_CPU 0.300 On_Core 2.399 IPC 1.479 Retire 0.410 (41.0%) FrontEnd 0.076 (7.6%) Spec 0.068 (6.8%) Backend 0.446 (44.6%) Elapsed 949.47 Procs 392 Maxrss 231K Minflt 40568424 Majflt 0 Inblock 0 Oublock 1072 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 27022 (54.4%) Nivcsw 22623 Utime 2229.506677 Stime 47.962057 Start 583656.30 Finish 584605.77
Now below the separate workloads:
sh - pid 31991 // blur On_CPU 0.311 On_Core 2.489 IPC 1.511 Retire 0.807 (80.7%) FrontEnd 0.169 (16.9%) Spec 0.139 (13.9%) Backend -0.115 (-11.5%) Elapsed 60.26 Procs 10 Maxrss 123K Minflt 4632988 Majflt 0 Inblock 0 Oublock 16 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 2248 (59.0%) Nivcsw 1563 Utime 145.145696 Stime 4.828832 Start 583659.06 Finish 583719.32 sh - pid 32030 // sharpen On_CPU 0.558 On_Core 4.464 IPC 1.599 Retire 0.899 (89.9%) FrontEnd 0.102 (10.2%) Spec 0.064 (6.4%) Backend -0.064 (-6.4%) Elapsed 60.37 Procs 10 Maxrss 84K Minflt 2471114 Majflt 0 Inblock 0 Oublock 16 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 1256 (36.4%) Nivcsw 2194 Utime 266.155163 Stime 3.322177 Start 583849.96 Finish 583910.33 sh - pid 32066 // resize On_CPU 0.328 On_Core 2.628 IPC 1.479 Retire 0.785 (78.5%) FrontEnd 0.121 (12.1%) Spec 0.162 (16.2%) Backend -0.067 (-6.7%) Elapsed 60.28 Procs 10 Maxrss 218K Minflt 2074168 Majflt 0 Inblock 0 Oublock 16 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 3203 (62.6%) Nivcsw 1915 Utime 155.858677 Stime 2.551043 Start 584041.29 Finish 584101.57 sh - pid 32101 // HWB color space On_CPU 0.252 On_Core 2.017 IPC 1.292 Retire 0.665 (66.5%) FrontEnd 0.169 (16.9%) Spec 0.232 (23.2%) Backend -0.067 (-6.7%) Elapsed 60.28 Procs 10 Maxrss 44K Minflt 2098220 Majflt 0 Inblock 0 Oublock 16 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 1631 (49.1%) Nivcsw 1693 Utime 118.890738 Stime 2.699423 Start 584231.78 Finish 584292.06 sh - pid 32164 On_CPU 0.125 On_Core 1.000 IPC 1.251 Retire 0.898 (89.8%) FrontEnd 0.279 (27.9%) Spec 0.178 (17.8%) Backend -0.355 (-35.5%) Elapsed 60.38 Procs 3 Maxrss 94K Minflt 2255789 Majflt 0 Inblock 0 Oublock 16 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 18 (19.4%) Nivcsw 75 Utime 58.004811 Stime 2.375055 Start 584422.54 Finish 584482.92
Metrics (AMD) - phoronix/graphics-magick
As a whole, the IPC for AMD is higher
phoronix-test-s - pid 32174 On_CPU 0.166 On_Core 2.653 IPC 1.795 FrontCyc 0.053 (5.3%) BackCyc 0.165 (16.5%) Elapsed 949.23 Procs 512 Maxrss 448K Minflt 49085597 Majflt 0 Inblock 0 Oublock 1048 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 76644 (24.2%) Nivcsw 240353 Utime 2447.632655 Stime 70.947529 Start 591984.20 Finish 592933.43
Here are the individual workloads:
sh - pid 32431 // blur On_CPU 0.155 On_Core 2.481 IPC 1.794 FrontCyc 0.105 (10.5%) BackCyc 0.192 (19.2%) Elapsed 60.18 Procs 18 Maxrss 125K Minflt 4840609 Majflt 0 Inblock 0 Oublock 16 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 7869 (35.1%) Nivcsw 14543 Utime 143.460783 Stime 5.874343 Start 591987.20 Finish 592047.38 sh - pid 32488 // sharpen On_CPU 0.358 On_Core 5.728 IPC 1.854 FrontCyc 0.016 (1.6%) BackCyc 0.131 (13.1%) Elapsed 60.21 Procs 18 Maxrss 85K Minflt 3314590 Majflt 0 Inblock 0 Oublock 16 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 4053 (11.1%) Nivcsw 32582 Utime 339.881471 Stime 4.973636 Start 592178.15 Finish 592238.36 sh - pid 32547 // resize On_CPU 0.175 On_Core 2.794 IPC 1.768 FrontCyc 0.088 (8.8%) BackCyc 0.162 (16.2%) Elapsed 60.16 Procs 18 Maxrss 383K Minflt 2861061 Majflt 0 Inblock 0 Oublock 16 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 6246 (27.2%) Nivcsw 16686 Utime 161.985703 Stime 6.113844 Start 592369.02 Finish 592429.18 sh - pid 32605 // HWB color space On_CPU 0.121 On_Core 1.936 IPC 1.649 FrontCyc 0.044 (4.4%) BackCyc 0.222 (22.2%) Elapsed 60.18 Procs 18 Maxrss 44K Minflt 2235478 Majflt 0 Inblock 0 Oublock 16 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 6569 (38.0%) Nivcsw 10704 Utime 113.145097 Stime 3.340982 Start 592559.69 Finish 592619.87 sh - pid 32661 // local adaptive thresholding On_CPU 0.062 On_Core 0.999 IPC 1.817 FrontCyc 0.049 (4.9%) BackCyc 0.188 (18.8%) Elapsed 60.31 Procs 3 Maxrss 94K Minflt 3113393 Majflt 0 Inblock 0 Oublock 16 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 18 (0.3%) Nivcsw 5489 Utime 57.080858 Stime 3.181444 Start 592750.28 Finish 592810.59
Process Tree - phoronix/graphics-magick
Process Tree
The process tree is fairly straightforward
31991) sh
31992) graphics-magick
31993) gm
31994) gm
31995) gm
31996) gm
31997) gm
31998) gm
31999) gm
32000) gm
Looking at the sum of all time scheduled on cores shows sharpen with the highest percentage (~55%) and adapter threshold with the lowest (~14%) and others closer to the 25-35%.
About this graph
Separating this amongst cores shows some chaos of individual cores sometimes scheduled for close to 100% and other much lower. Looks like the algorithms have sequential parts and parallel parts.
The IPC of these various parts also have some chaos but interestingly some common high+low modes in each.
About this graph
Topdown suggests for most the largest limiter are backend stalls with a moderate amount of bad speculation as well, particularly for the HWB color space application.
The overall topdown second-level metrics for the program:
retire 0.606 ms_uops 0.024 speculation 0.122 branch_misses 96.70% machine_clears 3.30% frontend 0.140 idq_uops_delivered_0 0.035 idq_uops_delivered_1 0.049 idq_uops_delivered_2 0.084 idq_uops_delivered_3 0.116 backend 0.133 resource_stalls.sb 0.014 stalls_ldm_pending 0.368
Moderately high speculation that appears to be branch misses. Not very high backend stalls and more likely memory reads.