gpu_sysman: Add ratio variant for temperature metric type

Signed-off-by: Eero Tamminen <eero.t.tamminen@intel.com>

diff --git a/src/gpu_sysman.c b/src/gpu_sysman.c
index c0f41ee37dc5520940a73ee4dbae1117a289ac25..a86107d8eb044c633954b7eee699a9e0d2f4309a 100644 (file)
--- a/src/gpu_sysman.c
+++ b/src/gpu_sysman.c
@@ -1469,14 +1469,19 @@ static bool gpu_temps(gpu_device_t *gpu) {
     gpu->temp_count = temp_count;
   }
 
-  metric_family_t fam = {
+  metric_family_t fam_temp = {
       .help = "Temperature sensor value (in Celsius) when queried",
       .name = METRIC_PREFIX "temperature_celsius",
       .type = METRIC_TYPE_GAUGE,
   };
+  metric_family_t fam_ratio = {
+      .help = "Temperature sensor value ratio to its max value when queried",
+      .name = METRIC_PREFIX "temperature_ratio",
+      .type = METRIC_TYPE_GAUGE,
+  };
   metric_t metric = {0};
 
-  bool ok = false;
+  bool reported_ratio = false, ok = false;
   for (i = 0; i < temp_count; i++) {
     zes_temp_properties_t props;
     if (zesTemperatureGetProperties(temps[i], &props) != ZE_RESULT_SUCCESS) {
@@ -1523,12 +1528,21 @@ static bool gpu_temps(gpu_device_t *gpu) {
     metric.value.gauge = value;
     metric_label_set(&metric, "location", type);
     metric_set_subdev(&metric, props.onSubdevice, props.subdeviceId);
-    metric_family_metric_append(&fam, metric);
+    metric_family_metric_append(&fam_temp, metric);
+
+    if (props.maxTemperature > 0 && (config.output & OUTPUT_RATIO)) {
+      metric.value.gauge = value / props.maxTemperature;
+      metric_family_metric_append(&fam_ratio, metric);
+      reported_ratio = true;
+    }
     ok = true;
   }
   if (ok) {
     metric_reset(&metric);
-    gpu_submit(gpu, &fam);
+    gpu_submit(gpu, &fam_temp);
+    if (reported_ratio) {
+      gpu_submit(gpu, &fam_ratio);
+    }
   }
   free(temps);
   return ok;

diff --git a/src/gpu_sysman_test.c b/src/gpu_sysman_test.c
index 325c0a5f5e7b753d2c07128ee24a45f741a72bb5..960f73bfc49b4d87a295a95663f584645f96b7c0 100644 (file)
--- a/src/gpu_sysman_test.c
+++ b/src/gpu_sysman_test.c
@@ -304,6 +304,7 @@ static ze_result_t metric_args_check(int callbit, const char *name,
 #define RAS_INIT 0
 #define RAS_INC 1
 
+#define TEMP_LIMIT 95
 #define TEMP_INIT 10
 #define TEMP_INC 5
 
@@ -393,7 +394,7 @@ ADD_METRIC(9, zesDeviceEnumPowerDomains, zes_pwr_handle_t,
            power_counter.energy += COUNTER_INC,
            power_counter.timestamp += TIME_INC)
 
-static zes_temp_properties_t temp_props;
+static zes_temp_properties_t temp_props = {.maxTemperature = TEMP_LIMIT};
 static double temperature = TEMP_INIT;
 static int dummy;
 
@@ -480,6 +481,9 @@ typedef struct {
   double last;
 } metrics_validation_t;
 
+#define TEMP_RATIO_INIT ((double)(TEMP_INIT) / (TEMP_LIMIT))
+#define TEMP_RATIO_INC ((double)(TEMP_INC) / (TEMP_LIMIT))
+
 #define MEM_RATIO_INIT ((double)MEMORY_INIT / MEMORY_SIZE)
 #define MEM_RATIO_INC ((double)MEMORY_INC / MEMORY_SIZE)
 
@@ -508,6 +512,7 @@ static metrics_validation_t valid_metrics[] = {
     {"memory_usage_ratio/HBM/system", false, false, MEM_RATIO_INIT,
      MEM_RATIO_INC, 0, 0.0},
     {"temperature_celsius", true, false, TEMP_INIT, TEMP_INC, 0, 0.0},
+    {"temperature_ratio", true, false, TEMP_RATIO_INIT, TEMP_RATIO_INC, 0, 0.0},
 
     /* while counters increase, per-time incremented value should stay same */
     {"energy_ujoules_total", true, false, COUNTER_START, COUNTER_INC, 0, 0.0},