Skip to contentSkip to conversation

perf jevents: Add dynamic IIO PCIe bandwidth metrics for Intel

PATCH In review Chun-Tse Shao 2 participants 4 messages Jul 9 – Sep 15 perf events

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf jevents: Add dynamic IIO PCIe bandwidth metrics for Intel

Generate per-IIO-device PCIe read and write bandwidth metrics (iio_bandwidth_read_per_device_0, etc.) grouped under iio_bandwidth_per_device for Sapphire Rapids and newer architectures in intel_metrics.py.

By using has_event() checks across up to 24 possible IIO stack indices, runtime perf automatically monitors only the IIO root complexes present on the system without failing on missing instances.

Assisted-by Gemini:gemini-3.1-pro-preview
---
 tools/perf/pmu-events/intel_metrics.py | 52 ++++++++++++++++++++++++++
 1 file changed, 52 insertions(+)
tools/perf/pmu-events/intel_metrics.py +52 -0
index bc2b920d3a0d..4f8a123a375d 100755
@@ -1132,6 +1132,57 @@ def UncoreUpiBw() -> Optional[MetricGroup]:
     ], description="UPI Bandwidth")
 
 
+def UncoreIioBw() -> Optional[MetricGroup]:
+    if _args.model not in [
+        "sapphirerapids",
+        "emeraldrapids",
+        "graniterapids",
+        "sierraforest",
+        "clearwaterforest",
+    ]:
+      return None
+
+    # On multi-socket x86 servers, there can be multiple IIO stacks per socket.
+    # We generate metrics up to index 23 with has_event check so runtime perf
+    # only measures the IIO instances present on the system.
+    max_iios = 24
+    metrics = []
+    scale = 4 / 1_000_000
+    for i in range(max_iios):
+      rd_event = Event(
+          f"uncore_iio_{i}/UNC_IIO_DATA_REQ_OF_CPU.MEM_READ.ALL_PARTS/"
+      )
+      wr_event = Event(
+          f"uncore_iio_{i}/UNC_IIO_DATA_REQ_OF_CPU.MEM_WRITE.ALL_PARTS/"
+      )
+      rd_bw = Select(d_ratio(rd_event, interval_sec), has_event(rd_event), 0)
+      wr_bw = Select(d_ratio(wr_event, interval_sec), has_event(wr_event), 0)
+      metrics.append(
+          MetricGroup(
+              f"iio_bandwidth_per_device_iio{i}",
+              [
+                  Metric(
+                      f"iio_bandwidth_read_per_device_{i}",
+                      f"IIO {i} read bandwidth",
+                      rd_bw,
+                      f"{scale}MB/s",
+                  ),
+                  Metric(
+                      f"iio_bandwidth_write_per_device_{i}",
+                      f"IIO {i} write bandwidth",
+                      wr_bw,
+                      f"{scale}MB/s",
+                  ),
+              ],
+          )
+      )
+    return MetricGroup(
+        "iio_bandwidth_per_device",
+        metrics,
+        description="IIO Read/Write Bandwidth per root complex",
+    )
+
+
 def main() -> None:
     global _args
 
@@ -1179,6 +1230,7 @@ def main() -> None:
         UncoreMemBw(),
         UncoreMemSat(),
         UncoreUpiBw(),
+        UncoreIioBw(),
     ])
 
     if _args.metricgroups:
Signature · 1 line
2.55.0.795.g602f6c329a-goog
#1 lore raw

+ Perry for reviewing the metric.

Quoted 89 lines · Chun-Tse Shao

Generate per-IIO-device PCIe read and write bandwidth metrics (iio_bandwidth_read_per_device_0, etc.) grouped under iio_bandwidth_per_device for Sapphire Rapids and newer architectures in intel_metrics.py.

By using has_event() checks across up to 24 possible IIO stack indices, runtime perf automatically monitors only the IIO root complexes present on the system without failing on missing instances.

Assisted-by Gemini:gemini-3.1-pro-preview
---
 tools/perf/pmu-events/intel_metrics.py | 52 ++++++++++++++++++++++++++
 1 file changed, 52 insertions(+)
tools/perf/pmu-events/intel_metrics.py +52 -1
index bc2b920d3a0d..4f8a123a375d 100755
@@ -1132,6 +1132,57 @@ def UncoreUpiBw() -> Optional[MetricGroup]:
     ], description="UPI Bandwidth")
 
 
+def UncoreIioBw() -> Optional[MetricGroup]:
+    if _args.model not in [
+        "sapphirerapids",
+        "emeraldrapids",
+        "graniterapids",
+        "sierraforest",
+        "clearwaterforest",
+    ]:
+      return None
+
+    # On multi-socket x86 servers, there can be multiple IIO stacks per socket.
+    # We generate metrics up to index 23 with has_event check so runtime perf
+    # only measures the IIO instances present on the system.
+    max_iios = 24
+    metrics = []
+    scale = 4 / 1_000_000
+    for i in range(max_iios):
+      rd_event = Event(
+          f"uncore_iio_{i}/UNC_IIO_DATA_REQ_OF_CPU.MEM_READ.ALL_PARTS/"
+      )
+      wr_event = Event(
+          f"uncore_iio_{i}/UNC_IIO_DATA_REQ_OF_CPU.MEM_WRITE.ALL_PARTS/"
+      )
+      rd_bw = Select(d_ratio(rd_event, interval_sec), has_event(rd_event), 0)
+      wr_bw = Select(d_ratio(wr_event, interval_sec), has_event(wr_event), 0)
+      metrics.append(
+          MetricGroup(
+              f"iio_bandwidth_per_device_iio{i}",
+              [
+                  Metric(
+                      f"iio_bandwidth_read_per_device_{i}",
+                      f"IIO {i} read bandwidth",
+                      rd_bw,
+                      f"{scale}MB/s",
+                  ),
+                  Metric(
+                      f"iio_bandwidth_write_per_device_{i}",
+                      f"IIO {i} write bandwidth",
+                      wr_bw,
+                      f"{scale}MB/s",
+                  ),
+              ],
+          )
+      )
+    return MetricGroup(
+        "iio_bandwidth_per_device",
+        metrics,
+        description="IIO Read/Write Bandwidth per root complex",
+    )
+
+
 def main() -> None:
     global _args
 
@@ -1179,6 +1230,7 @@ def main() -> None:
         UncoreMemBw(),
         UncoreMemSat(),
         UncoreUpiBw(),
+        UncoreIioBw(),
     ])
 
     if _args.metricgroups:
--

2.55.0.795.g602f6c329a-goog

reviewer Reviewed-by
#2 lore raw
Quoted 17 lines · Chun-Tse Shao

+ Perry for reviewing the metric.

Quoted 12 lines · Chun-Tse Shao

Generate per-IIO-device PCIe read and write bandwidth metrics (iio_bandwidth_read_per_device_0, etc.) grouped under iio_bandwidth_per_device for Sapphire Rapids and newer architectures in intel_metrics.py.

By using has_event() checks across up to 24 possible IIO stack indices, runtime perf automatically monitors only the IIO root complexes present on the system without failing on missing instances.

Assisted-by Gemini:gemini-3.1-pro-preview

Thanks,
Ian

Quoted 77 lines
Quoted 77 lines
---
 tools/perf/pmu-events/intel_metrics.py | 52 ++++++++++++++++++++++++++
 1 file changed, 52 insertions(+)
tools/perf/pmu-events/intel_metrics.py +52 -1
index bc2b920d3a0d..4f8a123a375d 100755
@@ -1132,6 +1132,57 @@ def UncoreUpiBw() -> Optional[MetricGroup]:
     ], description="UPI Bandwidth")
 
 
+def UncoreIioBw() -> Optional[MetricGroup]:
+    if _args.model not in [
+        "sapphirerapids",
+        "emeraldrapids",
+        "graniterapids",
+        "sierraforest",
+        "clearwaterforest",
+    ]:
+      return None
+
+    # On multi-socket x86 servers, there can be multiple IIO stacks per socket.
+    # We generate metrics up to index 23 with has_event check so runtime perf
+    # only measures the IIO instances present on the system.
+    max_iios = 24
+    metrics = []
+    scale = 4 / 1_000_000
+    for i in range(max_iios):
+      rd_event = Event(
+          f"uncore_iio_{i}/UNC_IIO_DATA_REQ_OF_CPU.MEM_READ.ALL_PARTS/"
+      )
+      wr_event = Event(
+          f"uncore_iio_{i}/UNC_IIO_DATA_REQ_OF_CPU.MEM_WRITE.ALL_PARTS/"
+      )
+      rd_bw = Select(d_ratio(rd_event, interval_sec), has_event(rd_event), 0)
+      wr_bw = Select(d_ratio(wr_event, interval_sec), has_event(wr_event), 0)
+      metrics.append(
+          MetricGroup(
+              f"iio_bandwidth_per_device_iio{i}",
+              [
+                  Metric(
+                      f"iio_bandwidth_read_per_device_{i}",
+                      f"IIO {i} read bandwidth",
+                      rd_bw,
+                      f"{scale}MB/s",
+                  ),
+                  Metric(
+                      f"iio_bandwidth_write_per_device_{i}",
+                      f"IIO {i} write bandwidth",
+                      wr_bw,
+                      f"{scale}MB/s",
+                  ),
+              ],
+          )
+      )
+    return MetricGroup(
+        "iio_bandwidth_per_device",
+        metrics,
+        description="IIO Read/Write Bandwidth per root complex",
+    )
+
+
 def main() -> None:
     global _args
 
@@ -1179,6 +1230,7 @@ def main() -> None:
         UncoreMemBw(),
         UncoreMemSat(),
         UncoreUpiBw(),
+        UncoreIioBw(),
     ])
 
     if _args.metricgroups:
--

2.55.0.795.g602f6c329a-goog

#3 lore raw
Quoted 17 lines · Ian Rogers
Quoted 15 lines · Chun-Tse Shao

+ Perry for reviewing the metric.

Quoted 10 lines · Chun-Tse Shao

Generate per-IIO-device PCIe read and write bandwidth metrics (iio_bandwidth_read_per_device_0, etc.) grouped under iio_bandwidth_per_device for Sapphire Rapids and newer architectures in intel_metrics.py.

By using has_event() checks across up to 24 possible IIO stack indices, runtime perf automatically monitors only the IIO root complexes present on the system without failing on missing instances.

Quoted 7 lines
Quoted 2 lines
Quoted 2 lines
Assisted-by Gemini:gemini-3.1-pro-preview

Thanks,
Ian

Thank you Ian for reviewing the patch! I had an offline discussion with Intel and we found:
1. iio_bandwidth_read/write is only missing on EMR and Intel fixed that through event updates:
lore.kernel.org/[email protected]
2. -A option does not work well with uncore events and I have a wip fix for that: lore.kernel.org/[email protected]

The patch is no longer necessary.

-CT

Quoted 78 lines
Quoted 78 lines
Quoted 78 lines
---
 tools/perf/pmu-events/intel_metrics.py | 52 ++++++++++++++++++++++++++
 1 file changed, 52 insertions(+)
tools/perf/pmu-events/intel_metrics.py +52 -1
index bc2b920d3a0d..4f8a123a375d 100755
@@ -1132,6 +1132,57 @@ def UncoreUpiBw() -> Optional[MetricGroup]:
     ], description="UPI Bandwidth")
 
 
+def UncoreIioBw() -> Optional[MetricGroup]:
+    if _args.model not in [
+        "sapphirerapids",
+        "emeraldrapids",
+        "graniterapids",
+        "sierraforest",
+        "clearwaterforest",
+    ]:
+      return None
+
+    # On multi-socket x86 servers, there can be multiple IIO stacks per socket.
+    # We generate metrics up to index 23 with has_event check so runtime perf
+    # only measures the IIO instances present on the system.
+    max_iios = 24
+    metrics = []
+    scale = 4 / 1_000_000
+    for i in range(max_iios):
+      rd_event = Event(
+          f"uncore_iio_{i}/UNC_IIO_DATA_REQ_OF_CPU.MEM_READ.ALL_PARTS/"
+      )
+      wr_event = Event(
+          f"uncore_iio_{i}/UNC_IIO_DATA_REQ_OF_CPU.MEM_WRITE.ALL_PARTS/"
+      )
+      rd_bw = Select(d_ratio(rd_event, interval_sec), has_event(rd_event), 0)
+      wr_bw = Select(d_ratio(wr_event, interval_sec), has_event(wr_event), 0)
+      metrics.append(
+          MetricGroup(
+              f"iio_bandwidth_per_device_iio{i}",
+              [
+                  Metric(
+                      f"iio_bandwidth_read_per_device_{i}",
+                      f"IIO {i} read bandwidth",
+                      rd_bw,
+                      f"{scale}MB/s",
+                  ),
+                  Metric(
+                      f"iio_bandwidth_write_per_device_{i}",
+                      f"IIO {i} write bandwidth",
+                      wr_bw,
+                      f"{scale}MB/s",
+                  ),
+              ],
+          )
+      )
+    return MetricGroup(
+        "iio_bandwidth_per_device",
+        metrics,
+        description="IIO Read/Write Bandwidth per root complex",
+    )
+
+
 def main() -> None:
     global _args
 
@@ -1179,6 +1230,7 @@ def main() -> None:
         UncoreMemBw(),
         UncoreMemSat(),
         UncoreUpiBw(),
+        UncoreIioBw(),
     ])
 
     if _args.metricgroups:
--

2.55.0.795.g602f6c329a-goog