In a toplev command with full tree, e.g.
$ toplev.py --no-desc -vl6 --nodes '+IPC,+Instructions,+Time,+SLOTS,+CLKS,+Mispredictions,+Big_Code,+Instruction_Fetch_BW,+Branching_Overhead,+DSB_Misses,+Memory_Bandwidth,+Memory_Latency,+Memory_Data_TLBs' -V m9b8IZ-x256-n8448-u01llv.toplev-vl6-perf.csv --frequency --metric-group +Summary
The events required by the Info.Bottleneck metrics should not be scheduled on their own. They are simply collected by the tree itself.
Try to fit them in same group isn't useful either since the smallest such metric require too many counters by itself.
As a result, toplev schedules at least 2x the needed events resulting in poor multiplexing rate as shown by this profile on ICX. Most nodes are sampled for [1.8%] of time!!
In a toplev command with full tree, e.g.
$ toplev.py --no-desc -vl6 --nodes '+IPC,+Instructions,+Time,+SLOTS,+CLKS,+Mispredictions,+Big_Code,+Instruction_Fetch_BW,+Branching_Overhead,+DSB_Misses,+Memory_Bandwidth,+Memory_Latency,+Memory_Data_TLBs' -V m9b8IZ-x256-n8448-u01llv.toplev-vl6-perf.csv --frequency --metric-group +Summary
The events required by the Info.Bottleneck metrics should not be scheduled on their own. They are simply collected by the tree itself. Try to fit them in same group isn't useful either since the smallest such metric require too many counters by itself. As a result, toplev schedules at least 2x the needed events resulting in poor multiplexing rate as shown by this profile on ICX. Most nodes are sampled for [1.8%] of time!!
This enhancement shall guide toplev to not schedule an event if the metrics Key column in the TMA-metrics spreadsheet start with "Info.Bot".