Skip to contentSkip to conversation

perf vendor events amd: Fix large metrics

PATCH Applied Sandipan Das 4 participants 5 messages Jul 6, 2023 – Jul 11, 2023 perf events

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 perf vendor events amd: Fix large metrics

There are cases where a metric requires more events than the number of available counters. E.g. AMD Zen, Zen 2 and Zen 3 processors have four data fabric counters but the "nps1_die_to_dram" metric has eight events. By default, the constituent events are placed in a group and since the events cannot be scheduled at the same time, the metric is not computed. The "all metrics" test also fails because of this.

Use the NO_GROUP_EVENTS constraint for such metrics which anyway expect the user to run perf with "--metric-no-group".

E.g.

  $ sudo perf test -v 101

Before:

101: perf all metrics test :
--- start ---
test child forked, pid 37131
Testing branch_misprediction_ratio
Testing all_remote_links_outbound
Testing nps1_die_to_dram
Metric 'nps1_die_to_dram' not printed in:
Error:
Invalid event (dram_channel_data_controller_4) in per-thread mode, enable system wide with '-a'. Testing macro_ops_dispatched
Testing all_l2_cache_accesses
Testing all_l2_cache_hits
Testing all_l2_cache_misses
Testing ic_fetch_miss_ratio
Testing l2_cache_accesses_from_l2_hwpf
Testing l2_cache_misses_from_l2_hwpf
Testing op_cache_fetch_miss_ratio
Testing l3_read_miss_latency
Testing l1_itlb_misses
test child finished with -1
---- end ----
perf all metrics test: FAILED!

After:

101: perf all metrics test :
--- start ---
test child forked, pid 43766
Testing branch_misprediction_ratio
Testing all_remote_links_outbound
Testing nps1_die_to_dram
Testing macro_ops_dispatched
Testing all_l2_cache_accesses
Testing all_l2_cache_hits
Testing all_l2_cache_misses
Testing ic_fetch_miss_ratio
Testing l2_cache_accesses_from_l2_hwpf
Testing l2_cache_misses_from_l2_hwpf
Testing op_cache_fetch_miss_ratio
Testing l3_read_miss_latency
Testing l1_itlb_misses
test child finished with 0
---- end ----
perf all metrics test: Ok

---

Previous versions can be found at:
v1: https://lore.kernel.org/all/[email protected]/

Changes in v2:
- As suggested by Ian, use the NO_GROUP_EVENTS constraint instead of retrying the test scenario with --metric-no-group.
- Change the commit message accordingly.

 tools/perf/pmu-events/arch/x86/amdzen1/recommended.json | 3 ++-
 tools/perf/pmu-events/arch/x86/amdzen2/recommended.json | 3 ++-
 tools/perf/pmu-events/arch/x86/amdzen3/recommended.json | 3 ++-
 3 files changed, 6 insertions(+), 3 deletions(-)
tools/perf/pmu-events/arch/x86/amdzen1/recommended.json +2 -1
index bf5083c1c260..4d28177325a0 100644
@@ -169,8 +169,9 @@
   },
   {
     "MetricName": "nps1_die_to_dram",
-    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die) (may need --metric-no-group)",
+    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die)",
     "MetricExpr": "dram_channel_data_controller_0 + dram_channel_data_controller_1 + dram_channel_data_controller_2 + dram_channel_data_controller_3 + dram_channel_data_controller_4 + dram_channel_data_controller_5 + dram_channel_data_controller_6 + dram_channel_data_controller_7",
+    "MetricConstraint": "NO_GROUP_EVENTS",
     "MetricGroup": "data_fabric",
     "PerPkg": "1",
     "ScaleUnit": "6.1e-5MiB"
tools/perf/pmu-events/arch/x86/amdzen2/recommended.json +2 -1
index a71694a043ba..60e19456d4c8 100644
@@ -169,8 +169,9 @@
   },
   {
     "MetricName": "nps1_die_to_dram",
-    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die) (may need --metric-no-group)",
+    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die)",
     "MetricExpr": "dram_channel_data_controller_0 + dram_channel_data_controller_1 + dram_channel_data_controller_2 + dram_channel_data_controller_3 + dram_channel_data_controller_4 + dram_channel_data_controller_5 + dram_channel_data_controller_6 + dram_channel_data_controller_7",
+    "MetricConstraint": "NO_GROUP_EVENTS",
     "MetricGroup": "data_fabric",
     "PerPkg": "1",
     "ScaleUnit": "6.1e-5MiB"
tools/perf/pmu-events/arch/x86/amdzen3/recommended.json +2 -1
index 988cf68ae825..3e9e1781812e 100644
@@ -205,10 +205,11 @@
   },
   {
     "MetricName": "nps1_die_to_dram",
-    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die) (may need --metric-no-group)",
+    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die)",
     "MetricExpr": "dram_channel_data_controller_0 + dram_channel_data_controller_1 + dram_channel_data_controller_2 + dram_channel_data_controller_3 + dram_channel_data_controller_4 + dram_channel_data_controller_5 + dram_channel_data_controller_6 + dram_channel_data_controller_7",
     "MetricGroup": "data_fabric",
     "PerPkg": "1",
+    "MetricConstraint": "NO_GROUP_EVENTS",
     "ScaleUnit": "6.1e-5MiB"
   }
 ]
Signature · 1 line
2.34.1
reviewer Acked-by on v2 1/1
#1 lore raw
Quoted 62 lines · Sandipan Das

There are cases where a metric requires more events than the number of available counters. E.g. AMD Zen, Zen 2 and Zen 3 processors have four data fabric counters but the "nps1_die_to_dram" metric has eight events. By default, the constituent events are placed in a group and since the events cannot be scheduled at the same time, the metric is not computed. The "all metrics" test also fails because of this.

Use the NO_GROUP_EVENTS constraint for such metrics which anyway expect the user to run perf with "--metric-no-group".

E.g.

  $ sudo perf test -v 101

Before:

101: perf all metrics test :
--- start ---
test child forked, pid 37131
Testing branch_misprediction_ratio
Testing all_remote_links_outbound
Testing nps1_die_to_dram
Metric 'nps1_die_to_dram' not printed in:
Error:
Invalid event (dram_channel_data_controller_4) in per-thread mode, enable system wide with '-a'. Testing macro_ops_dispatched
Testing all_l2_cache_accesses
Testing all_l2_cache_hits
Testing all_l2_cache_misses
Testing ic_fetch_miss_ratio
Testing l2_cache_accesses_from_l2_hwpf
Testing l2_cache_misses_from_l2_hwpf
Testing op_cache_fetch_miss_ratio
Testing l3_read_miss_latency
Testing l1_itlb_misses
test child finished with -1
---- end ----
perf all metrics test: FAILED!

After:

101: perf all metrics test :
--- start ---
test child forked, pid 43766
Testing branch_misprediction_ratio
Testing all_remote_links_outbound
Testing nps1_die_to_dram
Testing macro_ops_dispatched
Testing all_l2_cache_accesses
Testing all_l2_cache_hits
Testing all_l2_cache_misses
Testing ic_fetch_miss_ratio
Testing l2_cache_accesses_from_l2_hwpf
Testing l2_cache_misses_from_l2_hwpf
Testing op_cache_fetch_miss_ratio
Testing l3_read_miss_latency
Testing l1_itlb_misses
test child finished with 0
---- end ----
perf all metrics test: Ok

Quoted 3 lines

Will there be a PMU driver fix so that the perf_event_open fails for the group? That way the weak group would work.

Thanks,
Ian

Quoted 65 lines

---

Previous versions can be found at:
v1: https://lore.kernel.org/all/[email protected]/

Changes in v2:
- As suggested by Ian, use the NO_GROUP_EVENTS constraint instead of retrying the test scenario with --metric-no-group.
- Change the commit message accordingly.

 tools/perf/pmu-events/arch/x86/amdzen1/recommended.json | 3 ++-
 tools/perf/pmu-events/arch/x86/amdzen2/recommended.json | 3 ++-
 tools/perf/pmu-events/arch/x86/amdzen3/recommended.json | 3 ++-
 3 files changed, 6 insertions(+), 3 deletions(-)
tools/perf/pmu-events/arch/x86/amdzen1/recommended.json +2 -1
index bf5083c1c260..4d28177325a0 100644
@@ -169,8 +169,9 @@
   },
   {
     "MetricName": "nps1_die_to_dram",
-    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die) (may need --metric-no-group)",
+    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die)",
     "MetricExpr": "dram_channel_data_controller_0 + dram_channel_data_controller_1 + dram_channel_data_controller_2 + dram_channel_data_controller_3 + dram_channel_data_controller_4 + dram_channel_data_controller_5 + dram_channel_data_controller_6 + dram_channel_data_controller_7",
+    "MetricConstraint": "NO_GROUP_EVENTS",
     "MetricGroup": "data_fabric",
     "PerPkg": "1",
     "ScaleUnit": "6.1e-5MiB"
tools/perf/pmu-events/arch/x86/amdzen2/recommended.json +2 -1
index a71694a043ba..60e19456d4c8 100644
@@ -169,8 +169,9 @@
   },
   {
     "MetricName": "nps1_die_to_dram",
-    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die) (may need --metric-no-group)",
+    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die)",
     "MetricExpr": "dram_channel_data_controller_0 + dram_channel_data_controller_1 + dram_channel_data_controller_2 + dram_channel_data_controller_3 + dram_channel_data_controller_4 + dram_channel_data_controller_5 + dram_channel_data_controller_6 + dram_channel_data_controller_7",
+    "MetricConstraint": "NO_GROUP_EVENTS",
     "MetricGroup": "data_fabric",
     "PerPkg": "1",
     "ScaleUnit": "6.1e-5MiB"
tools/perf/pmu-events/arch/x86/amdzen3/recommended.json +2 -2
index 988cf68ae825..3e9e1781812e 100644
@@ -205,10 +205,11 @@
   },
   {
     "MetricName": "nps1_die_to_dram",
-    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die) (may need --metric-no-group)",
+    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die)",
     "MetricExpr": "dram_channel_data_controller_0 + dram_channel_data_controller_1 + dram_channel_data_controller_2 + dram_channel_data_controller_3 + dram_channel_data_controller_4 + dram_channel_data_controller_5 + dram_channel_data_controller_6 + dram_channel_data_controller_7",
     "MetricGroup": "data_fabric",
     "PerPkg": "1",
+    "MetricConstraint": "NO_GROUP_EVENTS",
     "ScaleUnit": "6.1e-5MiB"
   }
 ]
--

2.34.1

#2 lore raw

Hi Ian,

Quoted 63 lines · Ian Rogers
Quoted 62 lines · Sandipan Das

There are cases where a metric requires more events than the number of available counters. E.g. AMD Zen, Zen 2 and Zen 3 processors have four data fabric counters but the "nps1_die_to_dram" metric has eight events. By default, the constituent events are placed in a group and since the events cannot be scheduled at the same time, the metric is not computed. The "all metrics" test also fails because of this.

Use the NO_GROUP_EVENTS constraint for such metrics which anyway expect the user to run perf with "--metric-no-group".

E.g.

  $ sudo perf test -v 101

Before:

101: perf all metrics test :
--- start ---
test child forked, pid 37131
Testing branch_misprediction_ratio
Testing all_remote_links_outbound
Testing nps1_die_to_dram
Metric 'nps1_die_to_dram' not printed in:
Error:
Invalid event (dram_channel_data_controller_4) in per-thread mode, enable system wide with '-a'. Testing macro_ops_dispatched
Testing all_l2_cache_accesses
Testing all_l2_cache_hits
Testing all_l2_cache_misses
Testing ic_fetch_miss_ratio
Testing l2_cache_accesses_from_l2_hwpf
Testing l2_cache_misses_from_l2_hwpf
Testing op_cache_fetch_miss_ratio
Testing l3_read_miss_latency
Testing l1_itlb_misses
test child finished with -1
---- end ----
perf all metrics test: FAILED!

After:

101: perf all metrics test :
--- start ---
test child forked, pid 43766
Testing branch_misprediction_ratio
Testing all_remote_links_outbound
Testing nps1_die_to_dram
Testing macro_ops_dispatched
Testing all_l2_cache_accesses
Testing all_l2_cache_hits
Testing all_l2_cache_misses
Testing ic_fetch_miss_ratio
Testing l2_cache_accesses_from_l2_hwpf
Testing l2_cache_misses_from_l2_hwpf
Testing op_cache_fetch_miss_ratio
Testing l3_read_miss_latency
Testing l1_itlb_misses
test child finished with 0
---- end ----
perf all metrics test: Ok

Quoted 9 lines
Quoted 3 lines

Will there be a PMU driver fix so that the perf_event_open fails for the group? That way the weak group would work.

Yes, that's in our plan. Ravi (in CC) and I have discussed about adding group validation in the event_init() path.

- Sandipan

maintainer Applied
#3 lore raw
In reply to Ian Rogers

Em Thu, Jul 06, 2023 at 06:49:29AM -0700, Ian Rogers escreveu:

Quoted 63 lines · Sandipan Das
Quoted 62 lines

There are cases where a metric requires more events than the number of available counters. E.g. AMD Zen, Zen 2 and Zen 3 processors have four data fabric counters but the "nps1_die_to_dram" metric has eight events. By default, the constituent events are placed in a group and since the events cannot be scheduled at the same time, the metric is not computed. The "all metrics" test also fails because of this.

Use the NO_GROUP_EVENTS constraint for such metrics which anyway expect the user to run perf with "--metric-no-group".

E.g.

  $ sudo perf test -v 101

Before:

101: perf all metrics test :
--- start ---
test child forked, pid 37131
Testing branch_misprediction_ratio
Testing all_remote_links_outbound
Testing nps1_die_to_dram
Metric 'nps1_die_to_dram' not printed in:
Error:
Invalid event (dram_channel_data_controller_4) in per-thread mode, enable system wide with '-a'. Testing macro_ops_dispatched
Testing all_l2_cache_accesses
Testing all_l2_cache_hits
Testing all_l2_cache_misses
Testing ic_fetch_miss_ratio
Testing l2_cache_accesses_from_l2_hwpf
Testing l2_cache_misses_from_l2_hwpf
Testing op_cache_fetch_miss_ratio
Testing l3_read_miss_latency
Testing l1_itlb_misses
test child finished with -1
---- end ----
perf all metrics test: FAILED!

After:

101: perf all metrics test :
--- start ---
test child forked, pid 43766
Testing branch_misprediction_ratio
Testing all_remote_links_outbound
Testing nps1_die_to_dram
Testing macro_ops_dispatched
Testing all_l2_cache_accesses
Testing all_l2_cache_hits
Testing all_l2_cache_misses
Testing ic_fetch_miss_ratio
Testing l2_cache_accesses_from_l2_hwpf
Testing l2_cache_misses_from_l2_hwpf
Testing op_cache_fetch_miss_ratio
Testing l3_read_miss_latency
Testing l1_itlb_misses
test child finished with 0
---- end ----
perf all metrics test: Ok

Quoted 5 lines
Quoted 3 lines

Thanks, applied.

- Arnaldo

Quoted 71 lines

Will there be a PMU driver fix so that the perf_event_open fails for the group? That way the weak group would work.

Thanks,
Ian

Quoted 65 lines

---

Previous versions can be found at:
v1: https://lore.kernel.org/all/[email protected]/

Changes in v2:
- As suggested by Ian, use the NO_GROUP_EVENTS constraint instead of retrying the test scenario with --metric-no-group.
- Change the commit message accordingly.

 tools/perf/pmu-events/arch/x86/amdzen1/recommended.json | 3 ++-
 tools/perf/pmu-events/arch/x86/amdzen2/recommended.json | 3 ++-
 tools/perf/pmu-events/arch/x86/amdzen3/recommended.json | 3 ++-
 3 files changed, 6 insertions(+), 3 deletions(-)
tools/perf/pmu-events/arch/x86/amdzen1/recommended.json +2 -1
index bf5083c1c260..4d28177325a0 100644
@@ -169,8 +169,9 @@
   },
   {
     "MetricName": "nps1_die_to_dram",
-    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die) (may need --metric-no-group)",
+    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die)",
     "MetricExpr": "dram_channel_data_controller_0 + dram_channel_data_controller_1 + dram_channel_data_controller_2 + dram_channel_data_controller_3 + dram_channel_data_controller_4 + dram_channel_data_controller_5 + dram_channel_data_controller_6 + dram_channel_data_controller_7",
+    "MetricConstraint": "NO_GROUP_EVENTS",
     "MetricGroup": "data_fabric",
     "PerPkg": "1",
     "ScaleUnit": "6.1e-5MiB"
tools/perf/pmu-events/arch/x86/amdzen2/recommended.json +2 -1
index a71694a043ba..60e19456d4c8 100644
@@ -169,8 +169,9 @@
   },
   {
     "MetricName": "nps1_die_to_dram",
-    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die) (may need --metric-no-group)",
+    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die)",
     "MetricExpr": "dram_channel_data_controller_0 + dram_channel_data_controller_1 + dram_channel_data_controller_2 + dram_channel_data_controller_3 + dram_channel_data_controller_4 + dram_channel_data_controller_5 + dram_channel_data_controller_6 + dram_channel_data_controller_7",
+    "MetricConstraint": "NO_GROUP_EVENTS",
     "MetricGroup": "data_fabric",
     "PerPkg": "1",
     "ScaleUnit": "6.1e-5MiB"
tools/perf/pmu-events/arch/x86/amdzen3/recommended.json +2 -2
index 988cf68ae825..3e9e1781812e 100644
@@ -205,10 +205,11 @@
   },
   {
     "MetricName": "nps1_die_to_dram",
-    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die) (may need --metric-no-group)",
+    "BriefDescription": "Approximate: Combined DRAM B/bytes of all channels on a NPS1 node (die)",
     "MetricExpr": "dram_channel_data_controller_0 + dram_channel_data_controller_1 + dram_channel_data_controller_2 + dram_channel_data_controller_3 + dram_channel_data_controller_4 + dram_channel_data_controller_5 + dram_channel_data_controller_6 + dram_channel_data_controller_7",
     "MetricGroup": "data_fabric",
     "PerPkg": "1",
+    "MetricConstraint": "NO_GROUP_EVENTS",
     "ScaleUnit": "6.1e-5MiB"
   }
 ]
--

2.34.1

Signature · 1 line
- Arnaldo
maintainer Tested-by
#4 lore raw
Quoted 65 lines · Arnaldo Carvalho de Melo

Em Thu, Jul 06, 2023 at 06:49:29AM -0700, Ian Rogers escreveu:

Quoted 63 lines · Sandipan Das
Quoted 62 lines

There are cases where a metric requires more events than the number of available counters. E.g. AMD Zen, Zen 2 and Zen 3 processors have four data fabric counters but the "nps1_die_to_dram" metric has eight events. By default, the constituent events are placed in a group and since the events cannot be scheduled at the same time, the metric is not computed. The "all metrics" test also fails because of this.

Use the NO_GROUP_EVENTS constraint for such metrics which anyway expect the user to run perf with "--metric-no-group".

E.g.

  $ sudo perf test -v 101

Before:

101: perf all metrics test :
--- start ---
test child forked, pid 37131
Testing branch_misprediction_ratio
Testing all_remote_links_outbound
Testing nps1_die_to_dram
Metric 'nps1_die_to_dram' not printed in:
Error:
Invalid event (dram_channel_data_controller_4) in per-thread mode, enable system wide with '-a'. Testing macro_ops_dispatched
Testing all_l2_cache_accesses
Testing all_l2_cache_hits
Testing all_l2_cache_misses
Testing ic_fetch_miss_ratio
Testing l2_cache_accesses_from_l2_hwpf
Testing l2_cache_misses_from_l2_hwpf
Testing op_cache_fetch_miss_ratio
Testing l3_read_miss_latency
Testing l1_itlb_misses
test child finished with -1
---- end ----
perf all metrics test: FAILED!

After:

101: perf all metrics test :
--- start ---
test child forked, pid 43766
Testing branch_misprediction_ratio
Testing all_remote_links_outbound
Testing nps1_die_to_dram
Testing macro_ops_dispatched
Testing all_l2_cache_accesses
Testing all_l2_cache_hits
Testing all_l2_cache_misses
Testing ic_fetch_miss_ratio
Testing l2_cache_accesses_from_l2_hwpf
Testing l2_cache_misses_from_l2_hwpf
Testing op_cache_fetch_miss_ratio
Testing l3_read_miss_latency
Testing l1_itlb_misses
test child finished with 0
---- end ----
perf all metrics test: Ok

Quoted 7 lines
Quoted 5 lines
Quoted 3 lines

Thanks, applied.

If I'm not too late..

Thanks,
Namhyung