From f5b5681662e7c760c764dc5544e5f14149691b5f Mon Sep 17 00:00:00 2001
From: Mats Rynge <mats@rynge.net>
Date: Fri, 12 Dec 2025 12:18:24 -0800
Subject: [PATCH] OSPool: collect GPU usage data, mapped to capability and
 memory

---
 opensciencegrid/ospool-cm/prometheus.py | 82 +++++++++++++++++++++++++
 1 file changed, 82 insertions(+)

diff --git a/opensciencegrid/ospool-cm/prometheus.py b/opensciencegrid/ospool-cm/prometheus.py
index 1b06caad..f1031e4a 100755
--- a/opensciencegrid/ospool-cm/prometheus.py
+++ b/opensciencegrid/ospool-cm/prometheus.py
@@ -28,6 +28,16 @@
 ospool_claimed_gpus_count = Gauge("ospool_claimed_gpus_count", "Claimed GPUs", ["resource_name"])
 ospool_idle_gpus_count = Gauge("ospool_idle_gpus_count", "Idle GPUs", ["resource_name"])
 
+# gpus - capability
+ospool_total_gpus_capability = Gauge("ospool_total_gpus_capability", "Total GPUs", ["capability"])
+ospool_claimed_gpus_capability = Gauge("ospool_claimed_gpus_capability", "Claimed GPUs", ["capability"])
+ospool_idle_gpus_capability = Gauge("ospool_idle_gpus_capability", "Idle GPUs", ["capability"])
+
+# gpus - memory
+ospool_total_gpus_memory = Gauge("ospool_total_gpus_memory", "Total GPUs", ["memory"])
+ospool_claimed_gpus_memory = Gauge("ospool_claimed_gpus_memory", "Claimed GPUs", ["memory"])
+ospool_idle_gpus_memory = Gauge("ospool_idle_gpus_memory", "Idle GPUs", ["memory"])
+
 # pool wide attributes
 ospool_machine_ads = Gauge("ospool_machine_ads", "HTCondor machine ads")
 ospool_kernel_version = Gauge("ospool_kernel_versions", "Kernel versions in the pool", ["kernel_version"])
@@ -67,6 +77,18 @@ def entry_factory():
     return entry
 
 
+def entry_gpus():
+    '''
+    Make sure we have a base entry for given gpu capability
+    '''
+    entry = {
+        "total_gpus": 0,
+        "claimed_gpus": 0,
+        "idle_gpus": 0,
+    }
+    return entry
+
+
 def cm_resources_info(collector):
     '''
     Discover the sites, and then query for specifics per site.
@@ -138,6 +160,65 @@ def cm_resources_info(collector):
         ospool_idle_gpus_count.labels(resource).set(data["idle_gpus"])
 
 
+def cm_gpus(collector):
+    '''
+    Collect details on GPUs
+    '''
+
+    data = defaultdict(entry_gpus)
+
+    # capability iterate over all GPUs
+    ads = collector.query(ad_type=htcondor2.AdTypes.Startd,
+                          constraint="!isUndefined(GLIDEIN_ResourceName) && GPUs >= 1",
+                          projection=["GLIDEIN_ResourceName", "GPUs", "GPUs_Capability", "State"])
+    for ad in ads:
+        if "GPUs_Capability" not in ad:
+            ad["GPUs_Capability"] = "unknown"
+
+        d = data[ad["GPUs_Capability"]]
+
+        d["total_gpus"] += int(ad["GPUs"])
+        if ad["State"] != "Unclaimed":
+            d["claimed_gpus"] += int(ad["GPUs"])
+        d["idle_gpus"] = d["total_gpus"] - d["claimed_gpus"]
+
+    # done collecting data, update Prometheus
+    for capability, d in data.items():
+        ospool_total_gpus_capability.labels(capability).set(d["total_gpus"])
+        ospool_claimed_gpus_capability.labels(capability).set(d["claimed_gpus"])
+        ospool_idle_gpus_capability.labels(capability).set(d["idle_gpus"])
+
+    # now do the same iterating, but over memory
+    data = defaultdict(entry_gpus)
+    ads = collector.query(ad_type=htcondor2.AdTypes.Startd,
+                          constraint="!isUndefined(GLIDEIN_ResourceName) && GPUs >= 1",
+                          projection=["GLIDEIN_ResourceName", "GPUs", "GPUs_GlobalMemoryMb", "State"])
+    for ad in ads:
+        if "GPUs_GlobalMemoryMb" not in ad:
+            ad["GPUs_GlobalMemoryMb"] = 0
+
+        # convert to GB, value in to buckets
+        m = round((int(ad["GPUs_GlobalMemoryMb"]) / 1024))
+        for bucket in [0, 4, 8, 12, 16, 24, 32, 36, 40, 48, 64, 96, 128, 140, 192, 256, 512]:
+            if m <= bucket:
+                m = bucket
+                break
+
+        ad["GPUs_GlobalMemoryMb"] = str(m)
+
+        d = data[ad["GPUs_GlobalMemoryMb"]]
+
+        d["total_gpus"] += int(ad["GPUs"])
+        if ad["State"] != "Unclaimed":
+            d["claimed_gpus"] += int(ad["GPUs"])
+        d["idle_gpus"] = d["total_gpus"] - d["claimed_gpus"]
+
+    # done collecting data, update Prometheus
+    for memory, d in data.items():
+        ospool_total_gpus_memory.labels(memory).set(d["total_gpus"])
+        ospool_claimed_gpus_memory.labels(memory).set(d["claimed_gpus"])
+        ospool_idle_gpus_memory.labels(memory).set(d["idle_gpus"])
+
 def cm_pool_attributes(collector):
     '''
     Pool wide attributes
@@ -284,6 +365,7 @@ def cm_submitters_info(collector):
     while True:
         # can we determine if we are in ccb or cm mode?
         cm_resources_info(collector)
+        cm_gpus(collector)
         cm_pool_attributes(collector)
         cm_submitters_info(collector)
         time.sleep(20)