This test profile measures how long it takes to complete several reference GNU Octave files via octave-benchmark. GNU Octave is used for numerical computations and is an open-source alternative to MATLAB.
The benchmark runs six workloads in ~10 seconds. These tests are single-threaded and all testing was done pinned to core 1.
Description - phoronix/octave-benchmarksh - pid 17169 On_CPU 0.125 On_Core 0.999 IPC 1.555 Retire 0.323 (32.3%) FrontEnd 0.075 (7.5%) Spec 0.115 (11.5%) Backend 0.486 (48.6%) Elapsed 10.56 Procs 10 Maxrss 593K Minflt 1370931 Majflt 0 Inblock 0 Oublock 144 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 104 (70.3%) Nivcsw 44 Utime 9.525428 Stime 1.029286 Start 65087.73 Finish 65098.29
The test is On_CPU almost 100%. The IPC is moderately high with backend stalls as the largest issue.
Metrics (AMD) - phoronix/octave-benchmarksh - pid 2530 On_CPU 0.062 On_Core 0.999 IPC 1.476 FrontCyc 0.076 (7.6%) BackCyc 0.162 (16.2%) Elapsed 9.89 Procs 10 Maxrss 594K Minflt 1370690 Majflt 0 Inblock 0 Oublock 104 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 66 (6.3%) Nivcsw 982 Utime 8.835452 Stime 1.040407 Start 172659.53 Finish 172669.42
IPC on AMD is slightly lower.
Process Tree - phoronix/octave-benchmarkProcess Tree
17169) sh
17170) octave-benchmar
17171) octave-cli
17174) octave-cli
17175) octave-cli
17176) octave-cli
17177) octave-cli
17178) octave-cli
17179) head
17180) cut
The process tree is simple.
The workloads have some variation with the last workloads taking some system time and the first ones being predominately user time.
The IPC also varies some by workload. There is a one-second granularity of measurements, so not quite enough to sort out the six workloads from each other.
Topdown metrics show similar variations with occasional spikes in backend memory stalls and periods of speculative branch misses.
on_cpu 0.111 elapsed 37.064 utime 29.467 stime 3.397 nvcsw 1086 (89.98%) nivcsw 121 (10.02%) inblock 0 onblock 1192 retire 0.396 ms_uops 0.026 speculation 0.042 branch_misses 65.55% machine_clears 34.45% frontend 0.084 idq_uops_delivered_0 0.022 icache_stall 0.004 itlb_misses 0.002 idq_uops_delivered_1 0.034 idq_uops_delivered_2 0.047 idq_uops_delivered_3 0.062 dsb_ops 65.39% backend 0.479 resource_stalls.sb 0.069 stalls_ldm_pending 0.644 l2_refs 0.032 l2_misses 0.022 l2_miss_ratio 69.97% l3_refs 0.009 l3_misses 0.003 l3_miss_ratio 37.13%
Surprisingly high cache miss rates on L2/L3. Approximately 2/3 of the uops come from the uop cache. Overall speculation is low.