Re: [RFC PATCH 0/7] Share events between metrics

On Thu, May 7, 2020 at 2:47 PM Andi Kleen [off-list ref] wrote:

quoted

quoted

- without this change events within a metric may get scheduled
  together, after they may appear as part of a larger group and be
  multiplexed at different times, lowering accuracy - however, less
  multiplexing may compensate for this.

I agree the heuristic in this patch set is naive and would welcome to
improve it from your toplev experience. I think this change is
progress on TopDownL1 - would you agree?

TopdownL1 in non SMT mode should always fit. Inside a group
deduping always makes sense.

The problem is SMT mode where it doesn't fit. toplev tries
to group each node and each level together.

Thanks Andi, I've provided some examples of TopDownL3_SMT in the cover
letter of the v3 patch set:
https://lore.kernel.org/lkml/20200508053629.210324-1-irogers@google.com/ (local)
I tested sandybridge and cascadelake and the results look similar to
the non-SMT version. Let me know if there's a different variant to
test.

quoted

I'm wondering if what is needed are flags to control behavior. For
example, avoiding the use of groups altogether. For TopDownL1 I see.

Yes the current situation isn't great.

For Topdown your patch clearly is an improvement, I'm not sure
it's for everything though.

Probably the advanced heuristics are only useful for a few
formulas, most are very simple. So maybe it's ok. I guess
would need some testing over the existing formulas.

Agreed, do you have a pointer on a metric group where things would
obviously be worse? I started off with a cache miss and hit rate
metric and similar to topdown this approach is a benefit.

In v3 I've added a --metric-no-merge option to retain existing
grouping behavior, I've also added a --metric-no-group that avoids
groups for all metrics. This may be useful if the NMI watchdog can't
be disabled.

Thanks for the input!
Ian

-Andi