Fix Total CPU % on /workers tab to normalize by total nthreads (#9195)

Author: ernestprovo23

distributed/dashboard/components/scheduler.py

@@ -4076,6 +4076,7 @@ def __init__(self, scheduler, **kwargs):
             "host_disk_io.read_bps",
             "host_disk_io.write_bps",
             "cpu_fraction",
+            "_is_total",
         ]
         workers = self.scheduler.workers.values()
         self.extra_names = sorted(
@@ -4128,7 +4129,14 @@ def __init__(self, scheduler, **kwargs):
         }
 
         formatters = {
-            "cpu": NumberFormatter(format="0 %"),
+            # Use a pure number (0 to nthreads) on the total line and a %
+            # (e.g. 0 to 400% for 4 threads per worker ) on the individual workers.
+            # It would be very confusing to read e.g. 9000% on the total, whereas
+            # seeing that ~90 CPU equivalents are being fully used is more meaningful.
+            "cpu": HTMLTemplateFormatter(
+                template="<% if (_is_total) { %><%= (value).toFixed(1) %>"
+                "<% } else { %><%= Math.round(value * 100) %> %<% } %>"
+            ),
             "memory_percent": NumberFormatter(format="0.0 %"),
             "memory": NumberFormatter(format="0.0 b"),
             "memory_limit": NumberFormatter(format="0.0 b"),
@@ -4281,11 +4289,15 @@ def update(self):
             data["cpu"][-1] = ws.metrics["cpu"] / 100.0
             data["cpu_fraction"][-1] = ws.metrics["cpu"] / 100.0 / ws.nthreads
             data["nthreads"][-1] = ws.nthreads
+            data["_is_total"][-1] = False
 
         for name in self.names + self.extra_names:
             if name == "name":
                 data[name].insert(0, f"Total ({len(data[name])})")
                 continue
+            if name == "_is_total":
+                data[name].insert(0, True)
+                continue
             try:
                 if len(self.scheduler.workers) == 0:
                     total_data = None
@@ -4308,7 +4320,6 @@ def update(self):
                     total_data = (
                         sum(ws.metrics["cpu"] for ws in self.scheduler.workers.values())
                         / 100
-                        / len(self.scheduler.workers.values())
                     )
                 elif name == "cpu_fraction":
                     total_data = (

distributed/dashboard/tests/test_scheduler_bokeh.py

@@ -564,6 +564,17 @@ async def test_WorkerTable(c, s, a, b):
     assert all(nthreads)
     assert nthreads[0] == nthreads[1] + nthreads[2]
 
+    # Total CPU should show raw core count (sum of all worker CPU / 100)
+    cpu = wt.source.data["cpu"]
+    expected_cpu_total = sum(ws.metrics["cpu"] for ws in s.workers.values()) / 100
+    assert cpu[0] == expected_cpu_total
+
+    # _is_total flag should be set correctly
+    is_total = wt.source.data["_is_total"]
+    assert is_total[0] is True
+    assert is_total[1] is False
+    assert is_total[2] is False
+
 
 @gen_cluster(client=True)
 async def test_WorkerTable_custom_metrics(c, s, a, b):