This is a benchmark of BLAKE2 using the blake2s binary. BLAKE2 is a high-performance crypto alternative to MD5 and SHA-2/3.
The test is single-threaded. The overall runtime is fast and the test completes in less than 1/2 second. This means my graphs with granularity of 1 second don’t always capture everything.
Metrics (Intel) - phoronix/blake2sh - pid 8806 On_CPU 0.124 On_Core 0.992 IPC 1.405 Retire 0.280 (28.0%) FrontEnd 0.006 (0.6%) Spec 0.078 (7.8%) Backend 0.635 (63.5%) Elapsed 0.47 Procs 3 Maxrss 10K Minflt 238 Majflt 0 Inblock 0 Oublock 16 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 18 (78.3%) Nivcsw 5 Utime 0.466015 Stime 0.000000 Start 524403.13 Finish 524403.60
On_Core is almost 100%, the IPC is 1.41 and backend stalls seem to be the largest issue.
Metrics (AMD) - phoronix/blake2sh - pid 6948 On_CPU 0.062 On_Core 0.996 IPC 0.891 FrontCyc 0.004 (0.4%) BackCyc 0.003 (0.3%) Elapsed 0.77 Procs 3 Maxrss 10K Minflt 241 Majflt 0 Inblock 0 Oublock 16 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 18 (19.4%) Nivcsw 75 Utime 0.766571 Stime 0.000000 Start 407383.07 Finish 407383.84
IPC on my AMD system is considerably less.
Process Tree - phoronix/blake2
Process Tree
The process tree is simple
8806) sh elapsed=0.47 start=2.69 finish=3.16
8807) blake2 elapsed=0.46 start=2.70 finish=3.16
8808) blake2s elapsed=0.46 start=2.70 finish=3.16
A little too much noise to see the On_Core of 100%.
Similar noise with the IPC.
This is mostly the phoronix test suite running, but shows backend stalls.
on_cpu 0.041 elapsed 6.197 utime 1.817 stime 0.202 nvcsw 924 (97.67%) nivcsw 22 (2.33%) inblock 0 onblock 680 retire 0.358 ms_uops 0.015 speculation 0.021 branch_misses 51.68% machine_clears 48.32% frontend 0.077 idq_uops_delivered_0 0.035 icache_stall 0.009 itlb_misses 0.002 idq_uops_delivered_1 0.045 idq_uops_delivered_2 0.054 idq_uops_delivered_3 0.058 dsb_ops 80.40% backend 0.545 resource_stalls.sb 0.014 stalls_ldm_pending 0.602 l2_refs 0.006 l2_misses 0.003 l2_miss_ratio 61.13% l3_refs 0.002 l3_misses 0.000 l3_miss_ratio 16.17%
Backend stalls correlated with cache misses.
Next steps: Understand why IPC on AMD is considerably lower than Intel.