Add retirement latencies to perf events and other perf event and metric improvements #298
+57
−18
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Try to have fewer events classified as topic "other" preferring "cache" and "memory" topics. Remove some duplicated matches.
Use the mapfile.csv's event type of 'retirement latency' to locate a
retirement latency file. When create the perf event's dictionary add
retirement latency mean, min and max. For example on GNR in
cache.json:
```
...
{
"BriefDescription": "Retired load instructions with locked access.",
"Counter": "0,1,2,3",
"Data_LA": "1",
"EventCode": "0xd0",
"EventName": "MEM_INST_RETIRED.LOCK_LOADS",
"PublicDescription": "Counts retired load instructions with locked access.",
"RetirementLatencyMax": 5156,
"RetirementLatencyMean": 63.76,
"RetirementLatencyMin": 15,
"SampleAfterValue": "100007",
"UMask": "0x21"
},
...
```
Simplify metric thresholds to clean up unnecessary brackets in metrics parsed from TMA.
Without this Dependent_Loads_Weight isn't properly resolved.
edwarddavidbaker
approved these changes
Mar 24, 2025
There was a problem hiding this comment.
I generated the output from main and Ian's branched rebased onto the latest main.
Simplified metric threshold expressions.
--- main/alderlake/adl-metrics.json 2025-03-24 09:07:43.823351404 -0700
+++ retirement-latencies/alderlake/adl-metrics.json 2025-03-24 09:08:17.203371459 -0700
@@ -103,7 +103,7 @@
"MetricExpr": "tma_core_bound",
"MetricGroup": "TopdownL3;tma_L3_group;tma_core_bound_group",
"MetricName": "tma_allocation_restriction",
- "MetricThreshold": "(tma_allocation_restriction >0.10) & ((tma_core_bound >0.10) & ((tma_backend_bound >0.10)))",
+ "MetricThreshold": "tma_allocation_restriction > 0.1 & (tma_core_bound > 0.1 & tma_backend_bound > 0.1)",
"ScaleUnit": "100%",
"Unit": "cpu_atom"
},
@@ -113,7 +113,7 @@
"MetricExpr": "cpu_atom@TOPDOWN_BE_BOUND.ALL@ / (5 * cpu_atom@CPU_CLK_UNHALTED.CORE@)",
"MetricGroup": "Default;TopdownL1;tma_L1_group",
"MetricName": "tma_backend_bound",
- "MetricThreshold": "(tma_backend_bound >0.10)",
+ "MetricThreshold": "tma_backend_bound > 0.1",
"MetricgroupNoGroup": "TopdownL1;Default",
"PublicDescription": "Counts the total number of issue slots that were not consumed by the backend due to backend stalls. Note that uops must be available for consumption in order for this event to count. If a uop is not available (IQ is empty), this event will not count",
"ScaleUnit": "100%",Retired latency stats are added to events.
--- main/graniterapids/cache.json 2025-03-24 09:07:46.873356313 -0700
+++ retirement-latencies/graniterapids/cache.json 2025-03-24 09:08:20.623372472 -0700
@@ -344,60 +344,75 @@
"SampleAfterValue": "1000003",
"UMask": "0x83"
},
{
"BriefDescription": "Retired load instructions with locked access.",
"Counter": "0,1,2,3",
"Data_LA": "1",
"EventCode": "0xd0",
"EventName": "MEM_INST_RETIRED.LOCK_LOADS",
"PublicDescription": "Counts retired load instructions with locked access.",
+ "RetirementLatencyMax": 5156,
+ "RetirementLatencyMean": 63.76,
+ "RetirementLatencyMin": 15,
"SampleAfterValue": "100007",
"UMask": "0x21"
},
{
"BriefDescription": "Retired load instructions that split across a cacheline boundary.",
"Counter": "0,1,2,3",
"Data_LA": "1",
"EventCode": "0xd0",
"EventName": "MEM_INST_RETIRED.SPLIT_LOADS",
"PublicDescription": "Counts retired load instructions that split across a cacheline boundary.",
+ "RetirementLatencyMax": 4704,
+ "RetirementLatencyMean": 3.97,
+ "RetirementLatencyMin": 0,
"SampleAfterValue": "100003",
"UMask": "0x41"
},
|
@1perrytaylor Do you know why the last three
Discussed with Perry. This is expected. |
Use cpu_core rather than cpu PMU.
Ensure the correct spreadsheet and column name are used. This data isn't in mapfile.csv or the spreadsheet.
edwarddavidbaker
approved these changes
Mar 26, 2025
scripts/create_perf_json.py
Outdated
| @@ -1934,7 +1936,7 @@ def to_perf_json(self, outdir: Path): | |||
| perf_json.write('\n') | |||
|
|
|||
| # Skip hybrid because event grouping does not support it well yet | |||
| if self.shortname not in ['ADL', 'ADLN', 'ARL', 'LNL', 'MTL']: | |||
| if self.shortname not in ['ADL', 'ADLN', 'ARL', 'LNL', 'MTL', 'SRF']: | |||
There was a problem hiding this comment.
Please also add GRR If we need to skip grouping for E-core only products (these aren't hybrids).
Ensure the correct spreadsheet and column name are used. This data isn't in mapfile.csv or the spreadsheet.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Add retirement latencies to perf events on GNR.
Classify more cache and memory events topics.
Simplify metric thresholds.