Skip to contentSkip to conversation

perf metrics: Remove the "No_group" metric group

PATCH Applied Ian Rogers 4 participants 12 messages Apr 3, 2024 – Apr 8, 2024 perf events

lore
Share & tools
lore

Ian Rogers

author reviewer
#0 lore raw

PATCH 1/1 perf metrics: Remove the "No_group" metric group

Rather than place metrics without a metric group in "No_group" place them in a a metric group that is their name. Still allow such metrics to be selected if "No_group" is passed, this change just impacts perf list.

---
 tools/perf/util/metricgroup.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/util/metricgroup.c +2 -2
index 79ef6095ab28..6ec083af14a1 100644
@@ -455,7 +455,7 @@ static int metricgroup__add_to_mep_groups(const struct pmu_metric *pm,
 	const char *g;
 	char *omg, *mg;
 
-	mg = strdup(pm->metric_group ?: "No_group");
+	mg = strdup(pm->metric_group ?: pm->metric_name);
 	if (!mg)
 		return -ENOMEM;
 	omg = mg;
@@ -466,7 +466,7 @@ static int metricgroup__add_to_mep_groups(const struct pmu_metric *pm,
 		if (strlen(g))
 			me = mep_lookup(groups, g, pm->metric_name);
 		else
-			me = mep_lookup(groups, "No_group", pm->metric_name);
+			me = mep_lookup(groups, pm->metric_name, pm->metric_name);
 
 		if (me) {
 			me->metric_desc = pm->desc;
Signature · 1 line
2.44.0.478.gd926399ef9-goog
#1 lore raw
Quoted 4 lines · Ian Rogers

Rather than place metrics without a metric group in "No_group" place them in a a metric group that is their name. Still allow such metrics to be selected if "No_group" is passed, this change just impacts perf list.

So it looks like the "No_group" is not completely removed. They are just not seen in the perf list, but users can still use it via perf stat -M No_group, right?

If so, why we want to remove it from perf list? Where can the end user know which metrics are included in the No_group?

If the No_group is useless, why not completely remove it?

Thanks,
Kan

Quoted 28 lines
---
 tools/perf/util/metricgroup.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/util/metricgroup.c +2 -2
index 79ef6095ab28..6ec083af14a1 100644
@@ -455,7 +455,7 @@ static int metricgroup__add_to_mep_groups(const struct pmu_metric *pm,
 	const char *g;
 	char *omg, *mg;
 
-	mg = strdup(pm->metric_group ?: "No_group");
+	mg = strdup(pm->metric_group ?: pm->metric_name);
 	if (!mg)
 		return -ENOMEM;
 	omg = mg;
@@ -466,7 +466,7 @@ static int metricgroup__add_to_mep_groups(const struct pmu_metric *pm,
 		if (strlen(g))
 			me = mep_lookup(groups, g, pm->metric_name);
 		else
-			me = mep_lookup(groups, "No_group", pm->metric_name);
+			me = mep_lookup(groups, pm->metric_name, pm->metric_name);
 
 		if (me) {
 			me->metric_desc = pm->desc;
author reviewer
#2 lore raw
Quoted 9 lines · Kan
Quoted 4 lines · Ian Rogers

Rather than place metrics without a metric group in "No_group" place them in a a metric group that is their name. Still allow such metrics to be selected if "No_group" is passed, this change just impacts perf list.

Quoted 8 lines

So it looks like the "No_group" is not completely removed. They are just not seen in the perf list, but users can still use it via perf stat -M No_group, right?

If so, why we want to remove it from perf list? Where can the end user know which metrics are included in the No_group?

If the No_group is useless, why not completely remove it?

Agreed. For command line argument deprecation we usually keep the option but hide it from help with PARSE_OPT_HIDDEN, so I was trying to follow that pattern albeit that a metric group isn't a command line option it's an option to an option.

Thanks,
Ian

Quoted 31 lines

Thanks,
Kan

Quoted 28 lines
---
 tools/perf/util/metricgroup.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/util/metricgroup.c +2 -2
index 79ef6095ab28..6ec083af14a1 100644
@@ -455,7 +455,7 @@ static int metricgroup__add_to_mep_groups(const struct pmu_metric *pm,
      const char *g;
      char *omg, *mg;
 
-     mg = strdup(pm->metric_group ?: "No_group");
+     mg = strdup(pm->metric_group ?: pm->metric_name);
      if (!mg)
              return -ENOMEM;
      omg = mg;
@@ -466,7 +466,7 @@ static int metricgroup__add_to_mep_groups(const struct pmu_metric *pm,
              if (strlen(g))
                      me = mep_lookup(groups, g, pm->metric_name);
              else
-                     me = mep_lookup(groups, "No_group", pm->metric_name);
+                     me = mep_lookup(groups, pm->metric_name, pm->metric_name);
 
              if (me) {
                      me->metric_desc = pm->desc;
#3 lore raw
Quoted 17 lines · Ian Rogers
Quoted 16 lines · Kan
Quoted 4 lines · Ian Rogers

Rather than place metrics without a metric group in "No_group" place them in a a metric group that is their name. Still allow such metrics to be selected if "No_group" is passed, this change just impacts perf list.

So it looks like the "No_group" is not completely removed. They are just not seen in the perf list, but users can still use it via perf stat -M No_group, right?

If so, why we want to remove it from perf list? Where can the end user know which metrics are included in the No_group?

Quoted 7 lines
Quoted 1 line

If the No_group is useless, why not completely remove it?

Agreed. For command line argument deprecation we usually keep the option but hide it from help with PARSE_OPT_HIDDEN, so I was trying to follow that pattern albeit that a metric group isn't a command line option it's an option to an option.

Perf list has a deprecated option to show the deprecated events. The "No_group" should be a deprecated metrics group.

If so, to follow the same pattern, I think perf list should still display the "No_group" with the --deprecated option at least.

Thanks,
Kan

Quoted 34 lines

Thanks,
Ian

Quoted 31 lines

Thanks,
Kan

Quoted 28 lines
---
 tools/perf/util/metricgroup.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/util/metricgroup.c +2 -2
index 79ef6095ab28..6ec083af14a1 100644
@@ -455,7 +455,7 @@ static int metricgroup__add_to_mep_groups(const struct pmu_metric *pm,
      const char *g;
      char *omg, *mg;
 
-     mg = strdup(pm->metric_group ?: "No_group");
+     mg = strdup(pm->metric_group ?: pm->metric_name);
      if (!mg)
              return -ENOMEM;
      omg = mg;
@@ -466,7 +466,7 @@ static int metricgroup__add_to_mep_groups(const struct pmu_metric *pm,
              if (strlen(g))
                      me = mep_lookup(groups, g, pm->metric_name);
              else
-                     me = mep_lookup(groups, "No_group", pm->metric_name);
+                     me = mep_lookup(groups, pm->metric_name, pm->metric_name);
 
              if (me) {
                      me->metric_desc = pm->desc;
author reviewer
#4 lore raw
Quoted 29 lines · Kan
Quoted 17 lines · Ian Rogers
Quoted 16 lines · Kan
Quoted 4 lines · Ian Rogers

Rather than place metrics without a metric group in "No_group" place them in a a metric group that is their name. Still allow such metrics to be selected if "No_group" is passed, this change just impacts perf list.

So it looks like the "No_group" is not completely removed. They are just not seen in the perf list, but users can still use it via perf stat -M No_group, right?

If so, why we want to remove it from perf list? Where can the end user know which metrics are included in the No_group?

Quoted 7 lines
Quoted 1 line

If the No_group is useless, why not completely remove it?

Agreed. For command line argument deprecation we usually keep the option but hide it from help with PARSE_OPT_HIDDEN, so I was trying to follow that pattern albeit that a metric group isn't a command line option it's an option to an option.

Quoted 5 lines

Perf list has a deprecated option to show the deprecated events. The "No_group" should be a deprecated metrics group.

If so, to follow the same pattern, I think perf list should still display the "No_group" with the --deprecated option at least.

Such metrics would be shown twice, once under No_group and once under a metric group of their name. With deprecated events this isn't the case, you can only see them with --deprecated. Given we can see the metric without the No_group grouping, what is being added by having a No_group grouping? It feels entirely redundant and something we don't need to advertise.

Thanks,
Ian

Quoted 37 lines

Thanks,
Kan

Quoted 34 lines

Thanks,
Ian

Quoted 31 lines

Thanks,
Kan

Quoted 28 lines
---
 tools/perf/util/metricgroup.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/util/metricgroup.c +2 -2
index 79ef6095ab28..6ec083af14a1 100644
@@ -455,7 +455,7 @@ static int metricgroup__add_to_mep_groups(const struct pmu_metric *pm,
      const char *g;
      char *omg, *mg;
 
-     mg = strdup(pm->metric_group ?: "No_group");
+     mg = strdup(pm->metric_group ?: pm->metric_name);
      if (!mg)
              return -ENOMEM;
      omg = mg;
@@ -466,7 +466,7 @@ static int metricgroup__add_to_mep_groups(const struct pmu_metric *pm,
              if (strlen(g))
                      me = mep_lookup(groups, g, pm->metric_name);
              else
-                     me = mep_lookup(groups, "No_group", pm->metric_name);
+                     me = mep_lookup(groups, pm->metric_name, pm->metric_name);
 
              if (me) {
                      me->metric_desc = pm->desc;
#5 lore raw
Quoted 30 lines · Ian Rogers
Quoted 29 lines · Kan
Quoted 24 lines · Ian Rogers
Quoted 9 lines · Kan
Quoted 4 lines · Ian Rogers

Rather than place metrics without a metric group in "No_group" place them in a a metric group that is their name. Still allow such metrics to be selected if "No_group" is passed, this change just impacts perf list.

Quoted 8 lines

So it looks like the "No_group" is not completely removed. They are just not seen in the perf list, but users can still use it via perf stat -M No_group, right?

If so, why we want to remove it from perf list? Where can the end user know which metrics are included in the No_group?

If the No_group is useless, why not completely remove it?

Agreed. For command line argument deprecation we usually keep the option but hide it from help with PARSE_OPT_HIDDEN, so I was trying to follow that pattern albeit that a metric group isn't a command line option it's an option to an option.

Quoted 8 lines
Quoted 5 lines

Perf list has a deprecated option to show the deprecated events. The "No_group" should be a deprecated metrics group.

If so, to follow the same pattern, I think perf list should still display the "No_group" with the --deprecated option at least.

Such metrics would be shown twice, once under No_group and once under a metric group of their name.

You mean with the --deprecated option?
Yes, that's because the old/deprecated metrics group (No_group) is not complete removed. So both the new name and old/deprecated name are shown with the --deprecated option. The metrics which belong to both groups will be shown twice.

Without the --deprecated option, only the new group and its members are shown.

Quoted 5 lines

With deprecated events this isn't the
case, you can only see them with --deprecated. Given we can see the metric without the No_group grouping, what is being added by having a No_group grouping? It feels entirely redundant and something we don't need to advertise.

I just want to have a generic pattern for deprecating a metrics/metrics group that everybody can follow.

I treat the "No_group" as a normal metrics group name. So this patch is to introduce a new name, and hide the old name. Both new and old names can still be used.

If it's for a deprecated event, the expectation is to only see the new name by default, and see both new name and old name with the --deprecated option.

Now, if it's a generic deprecated metrics group, what's the expected behavior? I prefer to follow the same pattern as a deprecated event. If we do so, yes, there will be some redundancy with the --deprecated option, since some members may belong to both old and new groups. But I don't think it's an issue. It's normal that metrics belong to different groups.

Thanks,
Kan

author reviewer
#6 lore raw
Quoted 51 lines · Kan
Quoted 30 lines · Ian Rogers
Quoted 29 lines · Kan
Quoted 24 lines · Ian Rogers
Quoted 9 lines · Kan
Quoted 4 lines · Ian Rogers

Rather than place metrics without a metric group in "No_group" place them in a a metric group that is their name. Still allow such metrics to be selected if "No_group" is passed, this change just impacts perf list.

Quoted 8 lines

So it looks like the "No_group" is not completely removed. They are just not seen in the perf list, but users can still use it via perf stat -M No_group, right?

If so, why we want to remove it from perf list? Where can the end user know which metrics are included in the No_group?

If the No_group is useless, why not completely remove it?

Agreed. For command line argument deprecation we usually keep the option but hide it from help with PARSE_OPT_HIDDEN, so I was trying to follow that pattern albeit that a metric group isn't a command line option it's an option to an option.

Quoted 8 lines
Quoted 5 lines

Perf list has a deprecated option to show the deprecated events. The "No_group" should be a deprecated metrics group.

If so, to follow the same pattern, I think perf list should still display the "No_group" with the --deprecated option at least.

Such metrics would be shown twice, once under No_group and once under a metric group of their name.

You mean with the --deprecated option?
Yes, that's because the old/deprecated metrics group (No_group) is not complete removed. So both the new name and old/deprecated name are shown with the --deprecated option. The metrics which belong to both groups will be shown twice.

Without the --deprecated option, only the new group and its members are shown.

Quoted 8 lines
Quoted 5 lines

With deprecated events this isn't the
case, you can only see them with --deprecated. Given we can see the metric without the No_group grouping, what is being added by having a No_group grouping? It feels entirely redundant and something we don't need to advertise.

I just want to have a generic pattern for deprecating a metrics/metrics group that everybody can follow.

Currently there is no concept of a metric group in the json except for descriptions. Metrics and events share the same encoding, and the deprecated flag belongs to the event.

Quoted 3 lines

I treat the "No_group" as a normal metrics group name. So this patch is to introduce a new name, and hide the old name. Both new and old names can still be used.

Why are you using No_group? I stand firm that it has no real use.

Quoted 10 lines

If it's for a deprecated event, the expectation is to only see the new name by default, and see both new name and old name with the --deprecated option.

Now, if it's a generic deprecated metrics group, what's the expected behavior? I prefer to follow the same pattern as a deprecated event. If we do so, yes, there will be some redundancy with the --deprecated option, since some members may belong to both old and new groups. But I don't think it's an issue. It's normal that metrics belong to different groups.

What you are requesting here isn't something that is unreasonable, it is just something unconnected with this change and requires a reorganization of the json to facilitate. As such I consider it to be something for follow up work.

I think if we're going to restructure metric groups it would be nice to add a more tree-like structure which could be used to visualize metrics better. For example here:
https://lore.kernel.org/lkml/[email protected]/
the metrics could be shown under a tree like:
ldst
- ldst_total
- ldst_total_loads
- ldst_prcnt
- ldst_prcnt_loads
- ldst_prcnt_stores
- ldst_ret_lds
- ldst_ret_lds_1
- ldst_ret_lds_2
- ldst_ret_lds_3
- ldst_ret_sts
- ldst_ret_sts_1
- ldst_ret_sts_2
- ldst_ret_sts_3
- ldst_ld_hit_swpf
- ldst_atomic_lds

but again it is follow up work to do this. In this change I just wanted a way to list all sensibly grouped metrics or metrics in a group just on their own which doesn't require some kind of analysis of metric groups. No_group has no use so let's just get rid of it.

Thanks,
Ian

Quoted 2 lines

Thanks,
Kan

#7 lore raw
Quoted 100 lines · Ian Rogers
Quoted 51 lines · Kan
Quoted 30 lines · Ian Rogers
Quoted 29 lines · Kan
Quoted 24 lines · Ian Rogers
Quoted 9 lines · Kan
Quoted 4 lines · Ian Rogers

Rather than place metrics without a metric group in "No_group" place them in a a metric group that is their name. Still allow such metrics to be selected if "No_group" is passed, this change just impacts perf list.

Quoted 8 lines

So it looks like the "No_group" is not completely removed. They are just not seen in the perf list, but users can still use it via perf stat -M No_group, right?

If so, why we want to remove it from perf list? Where can the end user know which metrics are included in the No_group?

If the No_group is useless, why not completely remove it?

Agreed. For command line argument deprecation we usually keep the option but hide it from help with PARSE_OPT_HIDDEN, so I was trying to follow that pattern albeit that a metric group isn't a command line option it's an option to an option.

Quoted 8 lines
Quoted 5 lines

Perf list has a deprecated option to show the deprecated events. The "No_group" should be a deprecated metrics group.

If so, to follow the same pattern, I think perf list should still display the "No_group" with the --deprecated option at least.

Such metrics would be shown twice, once under No_group and once under a metric group of their name.

You mean with the --deprecated option?
Yes, that's because the old/deprecated metrics group (No_group) is not complete removed. So both the new name and old/deprecated name are shown with the --deprecated option. The metrics which belong to both groups will be shown twice.

Without the --deprecated option, only the new group and its members are shown.

Quoted 8 lines
Quoted 5 lines

With deprecated events this isn't the
case, you can only see them with --deprecated. Given we can see the metric without the No_group grouping, what is being added by having a No_group grouping? It feels entirely redundant and something we don't need to advertise.

I just want to have a generic pattern for deprecating a metrics/metrics group that everybody can follow.

Currently there is no concept of a metric group in the json except for descriptions. Metrics and events share the same encoding, and the deprecated flag belongs to the event.

Quoted 3 lines

I treat the "No_group" as a normal metrics group name. So this patch is to introduce a new name, and hide the old name. Both new and old names can still be used.

Why are you using No_group? I stand firm that it has no real use.

Quoted 10 lines

If it's for a deprecated event, the expectation is to only see the new name by default, and see both new name and old name with the --deprecated option.

Now, if it's a generic deprecated metrics group, what's the expected behavior? I prefer to follow the same pattern as a deprecated event. If we do so, yes, there will be some redundancy with the --deprecated option, since some members may belong to both old and new groups. But I don't think it's an issue. It's normal that metrics belong to different groups.

What you are requesting here isn't something that is unreasonable, it is just something unconnected with this change and requires a reorganization of the json to facilitate. As such I consider it to be something for follow up work.

I think if we're going to restructure metric groups it would be nice to add a more tree-like structure which could be used to visualize metrics better. For example here:
https://lore.kernel.org/lkml/[email protected]/
the metrics could be shown under a tree like:
ldst
- ldst_total
- ldst_total_loads
- ldst_prcnt
- ldst_prcnt_loads
- ldst_prcnt_stores
- ldst_ret_lds
- ldst_ret_lds_1

Quoted 9 lines

- ldst_ret_lds_2
- ldst_ret_lds_3
- ldst_ret_sts
- ldst_ret_sts_1
- ldst_ret_sts_2
- ldst_ret_sts_3
- ldst_ld_hit_swpf
- ldst_atomic_lds

Yes, a tree-like output looks much better.

Quoted 5 lines

but again it is follow up work to do this. In this change I just wanted a way to list all sensibly grouped metrics or metrics in a group just on their own which doesn't require some kind of analysis of metric groups. No_group has no use so let's just get rid of it.

I agree that there should be no one to use the No_group. Just hide it should be fine. Maybe we can have further discussion when someday we try to deprecate a meaningful metrics/metrics group.

Thanks,
Kan

#8 lore raw
In reply to Ian Rogers
Quoted 4 lines · Ian Rogers

Rather than place metrics without a metric group in "No_group" place them in a a metric group that is their name. Still allow such metrics to be selected if "No_group" is passed, this change just impacts perf list.

But what's the point of it? It will just make perf list more verbose, but I don't see any advantage.

-Andi

author reviewer
#9 lore raw
Quoted 9 lines · Andi Kleen
Quoted 4 lines · Ian Rogers

Rather than place metrics without a metric group in "No_group" place them in a a metric group that is their name. Still allow such metrics to be selected if "No_group" is passed, this change just impacts perf list.

But what's the point of it? It will just make perf list more verbose, but I don't see any advantage.

So it is possible to list all metrics, that's not changed here. The thing I'm looking to change is that when a metric is standalone it appears in "perf list metricgroups". The reason is that a metric group can gather a bunch of related metrics, say some form of read, write and total bandwidth, whereas something like an idle metric ("d_ratio(max(msr@tsc@ - msr@mperf@, 0), msr@tsc@)") that could get placed in No_group is more useful if it appears in a metric group of "idle". I'd put forward that nobody ever wants to run "idle" as part of "No_group" whereas being able to see it as a thing in metricgroups is useful. I want to be able to run "perf list metricgroups" and get groups of 1 or more metrics that someone might want to pass to "perf stat -M", currently this just shows when there is a group of more than 1 metric as there is no practice of putting a metric like "idle" into a metric group called "idle". We could update all metrics to make it so that when they don't have a metric group we add them to one with their name. We could do this in jevents.py. Those changes would make the No_group logic redundant, so we should remove it. Just updating the No_group logic in the perf command seemed like the minimal approach.

Thanks,
Ian

Quoted 1 line

-Andi

Reviewed-by on v1 1/1
#10 lore raw
In reply to Ian Rogers
Quoted 6 lines · Ian Rogers

Rather than place metrics without a metric group in "No_group" place them in a a metric group that is their name. Still allow such metrics to be selected if "No_group" is passed, this change just impacts perf list.

Thanks,
Kan

Quoted 26 lines
---
 tools/perf/util/metricgroup.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/util/metricgroup.c +2 -2
index 79ef6095ab28..6ec083af14a1 100644
@@ -455,7 +455,7 @@ static int metricgroup__add_to_mep_groups(const struct pmu_metric *pm,
 	const char *g;
 	char *omg, *mg;
 
-	mg = strdup(pm->metric_group ?: "No_group");
+	mg = strdup(pm->metric_group ?: pm->metric_name);
 	if (!mg)
 		return -ENOMEM;
 	omg = mg;
@@ -466,7 +466,7 @@ static int metricgroup__add_to_mep_groups(const struct pmu_metric *pm,
 		if (strlen(g))
 			me = mep_lookup(groups, g, pm->metric_name);
 		else
-			me = mep_lookup(groups, "No_group", pm->metric_name);
+			me = mep_lookup(groups, pm->metric_name, pm->metric_name);
 
 		if (me) {
 			me->metric_desc = pm->desc;
maintainer Applied
#11 lore raw
Quoted 11 lines · Kan
Quoted 6 lines · Ian Rogers

Rather than place metrics without a metric group in "No_group" place them in a a metric group that is their name. Still allow such metrics to be selected if "No_group" is passed, this change just impacts perf list.

Thanks, applied to perf-tools-next,

- Arnaldo