This is a standard video encoding performance test of Google’s libvpx library and the vpxenc command for the VP8/WebM format.
The test runs ./vpxenc with particular options on an image file. On both my 8-core Intel and 16-core AMD, this seems to run four processes in parallel.
Metrics (Intel) - phoronix/vpxencOn_CPU 0.329 On_Core 2.636 IPC 2.024 Retire 0.431 (43.1%) FrontEnd 0.111 (11.1%) Spec 0.158 (15.8%) Backend 0.299 (29.9%) Elapsed 259.33 Procs 6 Maxrss 289K Minflt 112781 Majflt 0 Inblock 0 Oublock 264 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 151733 (99.5%) Nivcsw 755 Utime 681.510345 Stime 2.034847 Start 100816.48 Finish 101075.81
The On_CPU is only 33%, so in addition to running only four threads, these are only kept busy on average 2/3 of the time. The percentage of voluntary context switches is ~99.5% of the total, so four threads with short bursts switching back and forth for short periods of time. Otherwise the IPC is moderately high with backend stalls being the largest issue.
Metrics (AMD) - phoronix/vpxencsh - pid 27234 On_CPU 0.168 On_Core 2.692 IPC 2.037 FrontCyc 0.000 (0.0%) BackCyc 0.000 (0.0%) Elapsed 278.21 Procs 6 Maxrss 289K Minflt 112784 Majflt 0 Inblock 0 Oublock 264 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 274866 (80.9%) Nivcsw 65034 Utime 747.130561 Stime 1.723176 Start 604864.48 Finish 605142.69
AMD metrics show an even lower On_CPU with only 4 threads running. Otherwise the IPC is similar.
Process Tree - phoronix/vpxenc
Process Tree
The process tree is simple.
25670) sh
25671) vpxenc
25672) vpxenc
25676) vpxenc
25677) vpxenc
25678) vpxenc
This results in a somewhat chaotic total of time spent scheduled.
About this graph
Even more chaotic if you look at individual threads.
These chaotic threads make the IPC tough to discern.
About this graph
However, the topdown metrics are more stable.
on_cpu 0.327 elapsed 779.596 utime 2033.577 stime 3.602 nvcsw 453038 (99.58%) nivcsw 1898 (0.42%) inblock 0 onblock 1456 retire 0.542 ms_uops 0.006 speculation 0.060 branch_misses 90.45% machine_clears 9.55% frontend 0.130 idq_uops_delivered_0 0.030 icache_stall 0.014 itlb_misses 0.000 idq_uops_delivered_1 0.046 idq_uops_delivered_2 0.070 idq_uops_delivered_3 0.113 dsb_ops 47.71% backend 0.268 resource_stalls.sb 0.015 stalls_ldm_pending 0.560 l2_refs 0.030 l2_misses 0.011 l2_miss_ratio 37.14% l3_refs 0.010 l3_misses 0.001 l3_miss_ratio 7.44%
The overall topdown metrics show the context switches, not many block of I/O, and backend stalls with miss ratios of 37% from L2 and 7% from L3. I expect performance might improve slightly with better thread pinning.