Metrics (Intel) - phoronix/ttsiod-rendererA portable GPL 3D software renderer that supports OpenMP and Intel Threading Building Blocks with many different rendering modes. This version does not use OpenGL but is entirely CPU/software based.
sh - pid 7311 On_CPU 0.975 On_Core 7.797 IPC 0.848 Retire 0.499 (49.9%) FrontEnd 0.214 (21.4%) Spec 0.062 (6.2%) Backend 0.225 (22.5%) Elapsed 31.80 Procs 10 Minflt 8497 Majflt 0 Utime 247.81 (100.0%) Stime 0.12 (0.0%) Start 398896.60 Finish 398928.40
The application runs for slightly over 30 seconds. The application is scheduled on the CPU just slightly less than 100% of the time. IPC is lower than some other benchmarks and suggests delays in backend are a factor.
Metrics (AMD) - phoronix/ttsiod-rendererttsiod-renderer - pid 4297 On_CPU 0.808 On_Core 12.930 IPC 1.053 FrontCyc 0.012 (1.2%) BackCyc 0.107 (10.7%) Elapsed 23.36 Procs 17 Minflt 8534 Majflt 0 Utime 300.94 (99.6%) Stime 1.11 (0.4%) Start 51139.78 Finish 51163.14
A case where the IPC on AMD is slightly higher than Intel.
Process Tree - phoronix/ttsiod-rendererProcess Tree
7311) sh elapsed=31.80 start=0.00 finish=31.80
7312) ttsiod-renderer elapsed=31.80 start=0.00 finish=31.80
7313) renderer elapsed=31.80 start=0.00 finish=31.80
7314) renderer elapsed=31.69 start=0.11 finish=31.80
7315) renderer elapsed=31.69 start=0.11 finish=31.80
7316) renderer elapsed=31.69 start=0.11 finish=31.80
7317) renderer elapsed=31.69 start=0.11 finish=31.80
7318) renderer elapsed=31.69 start=0.11 finish=31.80
7319) renderer elapsed=31.69 start=0.11 finish=31.80
7320) renderer elapsed=31.69 start=0.11 finish=31.80
The process tree is similar to others with parallel activities started on cores. One subtle difference is there are only seven children, not eight, a potential factor in why On_CPU isn’t quite 100%.
Resource metrics show some, but not abnormally high voluntary context switches (that would indicate processes waiting on resources), also a reasonable amount of blocks read in for a 30 second workload, so having a faster disk might help.
utime: 742.795869 stime: 0.371598 maxrss: 28K minflt: 71799 majflt: 2 nswap: 0 inblock: 5880 oublock: 680 msgsnd: 0 msgrcv: 0 nsignals: 0 nvcsw: 31249 nivcsw: 5130
Add up the total usage of all CPUs and it looks evenly allocated just slightly less than 100%.
About this graph
However, plot the cores separately and one sees how different cores become idle for short periods. This behavior is where I might expect the OS scheduler policy to have an effect.
IPC across the three runs of the workload.
About this graph
Backend stalls are slightly higher than average.
Next steps: Drill down on factors influencing backend stalls such as cache and memory.