Metrics (Intel) - phoronix/compress-pbzip2This test measures the time needed to compress a file (a .tar package of the Linux kernel source code) using BZIP2 compression.
sh - pid 2426 On_CPU 0.956 On_Core 7.645 IPC 0.913 Retire 0.451 (45.1%) FrontEnd 0.098 (9.8%) Spec 0.300 (30.0%) Backend 0.151 (15.1%) Elapsed 9.71 Procs 14 Maxrss 740K Minflt 191115 Majflt 0 Inblock 0 Oublock 0 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 212 (21.7%) Nivcsw 763 Utime 73.876767 Stime 0.357475 Start 743780.58 Finish 743790.29
The voluntary context switches and On_CPU of 95% suggest some I/O (despite the Inblock/Oublock) but overall, this is scheduled mostly on the processor. The speculative miss rate is very high and the IPC slightly on the low side.
Metrics (AMD) - phoronix/compress-pbzip2sh - pid 6298 On_CPU 0.933 On_Core 14.931 IPC 0.928 FrontCyc 0.130 (13.0%) BackCyc 0.129 (12.9%) Elapsed 5.35 Procs 22 Maxrss 763K Minflt 198517 Majflt 0 Inblock 0 Oublock 0 Msgsnd 0 Msgrcv 0 Nsignals 0 Nvcsw 361 (4.3%) Nivcsw 8078 Utime 79.398898 Stime 0.480230 Start 753843.64 Finish 753848.99
Similar IPC between AMD and Intel.
Process Tree - phoronix/compress-pbzip2
Process Tree
The process tree is simple.
2426) sh
2427) compress-pbzip2
2428) pbzip2
2429) pbzip2
2430) pbzip2
2431) pbzip2
2432) pbzip2
2433) pbzip2
2434) pbzip2
2435) pbzip2
2436) pbzip2
2437) pbzip2
2438) pbzip2
2439) pbzip2
About this graph
Shows all cores scheduled most of the time.
The IPC is somewhat lower and has some noise.
About this graph
There is a high amount of speculative misses.
retire 0.443 ms_uops 0.002 speculation 0.296 branch_misses 81.24% machine_clears 18.76% frontend 0.100 idq_uops_delivered_0 0.023 idq_uops_delivered_1 0.037 idq_uops_delivered_2 0.054 idq_uops_delivered_3 0.085 backend 0.162 resource_stalls.sb 0.030 stalls_ldm_pending 0.236
Most of the bad speculation is due to branch prediction, but there are also machine clears.
Next steps: Understand why the branch prediction is poor.