On Thu, May 7, 2020 at 10:48 AM Andi Kleen <a...@linux.intel.com> wrote:
>
> On Thu, May 07, 2020 at 01:14:29AM -0700, Ian Rogers wrote:
> > Metric groups contain metrics. Metrics create groups of events to
> > ideally be scheduled together. Often metrics refer to the same events,
> > for example, a cache hit and cache miss rate. Using separate event
> > groups means these metrics are multiplexed at different times and the
> > counts don't sum to 100%. More multiplexing also decreases the
> > accuracy of the measurement.
> >
> > This change orders metrics from groups or the command line, so that
> > the ones with the most events are set up first. Later metrics see if
> > groups already provide their events, and reuse them if
> > possible. Unnecessary events and groups are eliminated.
>
> Note this actually may make multiplexing errors worse.
>
> For metrics it is often important that all the input values to
> the metric run at the same time.
>
> So e.g. if you have two metrics and they each fit into a group,
> but not together, even though you have more multiplexing it
> will give more accurate results for each metric.
>
> I think you change can make sense for metrics that don't fit
> into single groups anyways. perf currently doesn't quite know
> this but some heuristic could be added.
>
> But I wouldn't do it for simple metrics that fit into groups.
> The result may well be worse.
>
> My toplev tool has some heuristics for this, also some more
> sophisticated ones that tracks subexpressions. That would
> be far too complicated for perf likely.
>
> -Andi
Thanks Andi!
I was trying to be mindful of the multiplexing issue in the description:
> - 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?
I'm wondering if what is needed are flags to control behavior. For
example, avoiding the use of groups altogether. For TopDownL1 I see.
Currently:
27,294,614,172 idq_uops_not_delivered.core # 0.3
Frontend_Bound (49.96%)
24,498,363,923 cycles
(49.96%)
21,449,143,854 uops_issued.any # 0.1
Bad_Speculation (16.68%)
16,450,676,961 uops_retired.retire_slots
(16.68%)
880,423,103 int_misc.recovery_cycles
(16.68%)
24,180,775,568 cycles
(16.68%)
27,662,201,567 idq_uops_not_delivered.core # 0.5
Backend_Bound (16.67%)
25,354,331,290 cycles
(16.67%)
22,642,218,398 uops_issued.any
(16.67%)
17,564,211,383 uops_retired.retire_slots
(16.67%)
896,938,527 int_misc.recovery_cycles
(16.67%)
17,872,454,517 uops_retired.retire_slots # 0.2 Retiring
(16.68%)
25,122,100,836 cycles
(16.68%)
15,101,167,608 inst_retired.any # 0.6 IPC
(33.34%)
24,977,816,793 cpu_clk_unhalted.thread
(33.34%)
24,868,717,684 cycles
# 99474870736.0
SLOTS (49.98%)
With proposed (RFC) sharing of events over metrics:
22,780,823,620 cycles
# 91123294480.0
SLOTS
# 0.2 Retiring
# 0.3
Frontend_Bound
# 0.1
Bad_Speculation
# 0.4
Backend_Bound (50.01%)
26,097,362,439 idq_uops_not_delivered.core
(50.01%)
790,521,504 int_misc.recovery_cycles
(50.01%)
21,025,308,329 uops_issued.any
(50.01%)
17,041,506,149 uops_retired.retire_slots
(50.01%)
22,964,891,526 cpu_clk_unhalted.thread # 0.6 IPC
(49.99%)
14,531,724,741 inst_retired.any
(49.99%)
No groups:
1,517,455,258 cycles
# 6069821032.0 SLOTS
# 0.1 Retiring
# 0.3
Frontend_Bound
# 0.1
Bad_Speculation
# 0.5
Backend_Bound (85.64%)
1,943,047,724 idq_uops_not_delivered.core
(85.61%)
54,257,713 int_misc.recovery_cycles
(85.63%)
1,050,787,137 uops_issued.any
(85.63%)
881,310,530 uops_retired.retire_slots
(85.68%)
1,553,561,836 cpu_clk_unhalted.thread # 0.5 IPC
(71.81%)
742,087,439 inst_retired.any
(85.85%)
So with no groups there is a lot less multiplexing.
So I'm thinking of two flags:
- disable sharing of events between metrics - defaulted off - this
keeps the current behavior in case there is a use-case where
multiplexing is detrimental. I'm not sure how necessary this flag is,
if we could quantify it based on experience elsewhere it'd be nice.
Default off as without sharing metrics within a metric group fail to
add to 100%. Fwiw, I can imagine phony metrics that exist just to
cause sharing of events within a group.
- disable grouping of events in metrics - defaulted off - this would
change the behavior of groups like TopDownL1 as I show above for "no
groups".
I see in toplev:
https://github.com/andikleen/pmu-tools/wiki/toplev-manual
--no-group which is similar to the second flag.
Do you have any pointers in toplev for better grouping heuristics?
Thoughts and better ways to do this very much appreciated! Thanks,
Ian
