On 06/02/25 12:09, Sebastian Andrzej Siewior wrote: > On 2025-02-06 12:02:49 [-0300], Rafael Folco wrote: >> On 06/02/25 09:00, Sebastian Andrzej Siewior wrote: >>> On 2025-02-06 06:16:22 [-0300], Rafael Folco wrote: >>>> Make min/max consistent with avg by recording the latency samples >>>> as float numbers instead of integers. >>> >>> Why? >> >> - More precise numbers of Max/Min for extreme low latency cases >> - Distinguish zero latency in bucket 001 (us) >> - Distinguish values from upper/lower boundaries in a bucket >> - Clarify confusion on average (round up) vs maximum in the next bucket (casting) >> >> Instead of: >> Core: 3 4 5 >> Counter Freq: 2100 2100 2100 (MHz) >> 001 (us): 14880514 14880447 14880442 >> 002 (us): 0 0 0 >> 003 (us): 0 0 0 >> 004 (us): 0 0 0 >> 005 (us): 0 0 0 (including overflows) >> Minimum: 0 0 0 (us) >> Average: 1.000 1.000 1.000 (us) >> Maximum: 0 0 0 (us) >> Max-Min: 0 0 0 (us) >> >> We'd see: >> Core: 3 4 5 >> Counter Freq: 2100 2100 2100 (MHz) >> 001 (us): 12578842 12491108 12578752 >> 002 (us): 0 0 0 >> 003 (us): 0 0 0 >> 004 (us): 0 0 0 >> 005 (us): 0 0 0 (including overflows) >> Minimum: 0.033 0.035 0.035 (us) >> Average: 1.000 1.000 1.000 (us) >> Maximum: 0.129 0.153 0.134 (us) >> Max-Min: 0.095 0.118 0.099 (us) > > so core 3 has min 33ns and max 129ns while 1us on average? Yes, because avg rounds up. There are some other potential fixes coming like: - stop rounding up avg - splitting bucket 001 (us) which is bigger (0 - 1.99999) - adding the max bucket to the end (do not limit to 32us) ... but these are out of scope for this change. Folco > > Sebastian >
