Autoscale by gpu signals too

Just take max of cpu and gpu load.

2 files changed, 65 insertions, 4 deletions

diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/MetricsResponse.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/MetricsResponse.java
index 8b7a2bafc40..fa9cfff0d68 100644
--- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/MetricsResponse.java
+++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/MetricsResponse.java
@@ -120,11 +120,13 @@ public class MetricsResponse {
         cpu { // a node resource
 
             @Override
-            public List<String> metricResponseNames() { return List.of(HostedNodeAdminMetrics.CPU_UTIL.baseName()); }
+            public List<String> metricResponseNames() {
+                return List.of(HostedNodeAdminMetrics.CPU_UTIL.baseName(), HostedNodeAdminMetrics.GPU_UTIL.baseName());
+            }
 
             @Override
             double computeFinal(ListMap<String, Double> values) {
-                return values.values().stream().flatMap(List::stream).mapToDouble(v -> v).average().orElse(0) / 100; // % to ratio
+                return values.values().stream().flatMap(List::stream).mapToDouble(v -> v).max().orElse(0) / 100; // % to ratio
             }
 
         },
@@ -133,11 +135,17 @@ public class MetricsResponse {
             @Override
             public List<String> metricResponseNames() {
                 return List.of(HostedNodeAdminMetrics.MEM_UTIL.baseName(),
-                               SearchNodeMetrics.CONTENT_PROTON_RESOURCE_USAGE_MEMORY.average());
+                               SearchNodeMetrics.CONTENT_PROTON_RESOURCE_USAGE_MEMORY.average(),
+                               HostedNodeAdminMetrics.GPU_MEM_USED.baseName(),
+                               HostedNodeAdminMetrics.GPU_MEM_TOTAL.baseName());
             }
 
             @Override
             double computeFinal(ListMap<String, Double> values) {
+                return Math.max(gpuMemUtil(values), cpuMemUtil(values));
+            }
+
+            private double cpuMemUtil(ListMap<String, Double> values) {
                 var valueList = values.get(SearchNodeMetrics.CONTENT_PROTON_RESOURCE_USAGE_MEMORY.average()); // prefer over mem.util
                 if ( ! valueList.isEmpty()) return valueList.get(0);
 
@@ -147,6 +155,12 @@ public class MetricsResponse {
                 return 0;
             }
 
+            private double gpuMemUtil(ListMap<String, Double> values) {
+                var usedGpuMemory = values.get(HostedNodeAdminMetrics.GPU_MEM_USED.baseName()).stream().mapToDouble(v -> v).sum();
+                var totalGpuMemory = values.get(HostedNodeAdminMetrics.GPU_MEM_TOTAL.baseName()).stream().mapToDouble(v -> v).sum();
+                return totalGpuMemory > 0 ? usedGpuMemory / totalGpuMemory : 0;
+            }
+
         },
         disk { // a node resource
 
diff --git a/node-repository/src/test/java/com/yahoo/vespa/hosted/provision/autoscale/MetricsV2MetricsFetcherTest.java b/node-repository/src/test/java/com/yahoo/vespa/hosted/provision/autoscale/MetricsV2MetricsFetcherTest.java
index 24697d02681..01a4e96a195 100644
--- a/node-repository/src/test/java/com/yahoo/vespa/hosted/provision/autoscale/MetricsV2MetricsFetcherTest.java
+++ b/node-repository/src/test/java/com/yahoo/vespa/hosted/provision/autoscale/MetricsV2MetricsFetcherTest.java
@@ -80,7 +80,7 @@ public class MetricsV2MetricsFetcherTest {
             assertTrue(values.get(0).getSecond().stable());
         }
 
-        {
+        { // read response 2 when unstable
             httpClient.cannedResponse = cannedResponseForApplication2;
             try (Mutex lock = tester.nodeRepository().applications().lock(application1)) {
                 tester.nodeRepository().nodes().write(tester.nodeRepository().nodes().list(Node.State.active).owner(application2)
@@ -89,6 +89,18 @@ public class MetricsV2MetricsFetcherTest {
             List<Pair<String, NodeMetricSnapshot>> values = new ArrayList<>(fetcher.fetchMetrics(application2).get().nodeMetrics());
             assertFalse(values.get(0).getSecond().stable());
         }
+
+        {
+            httpClient.cannedResponse = cannedResponseForApplication3;
+            List<Pair<String, NodeMetricSnapshot>> values = new ArrayList<>(fetcher.fetchMetrics(application2).get().nodeMetrics());
+            assertEquals("http://host-3.yahoo.com:4080/metrics/v2/values?consumer=autoscaling",
+                         httpClient.requestsReceived.get(1));
+            assertEquals(1, values.size());
+            assertEquals("host-3.yahoo.com", values.get(0).getFirst());
+            assertEquals(0.13, values.get(0).getSecond().load().cpu(), delta);
+            assertEquals(0.9375, values.get(0).getSecond().load().memory(), delta);
+        }
+
     }
 
     private static class MockHttpClient implements MetricsV2MetricsFetcher.AsyncHttpClient {
@@ -208,7 +220,42 @@ public class MetricsV2MetricsFetcherTest {
                               {
                                 "values": {
                                   "cpu.util": 10,
+                                  "gpu.util": 8,
+                                  "mem.util": 15,
+                                  "gpu.memory.used": 0,
+                                  "gpu.memory.total": 8,
+                                  "disk.util": 20,
+                                  "application_generation.last": 3,
+                                  "in_service.last": 0
+                                },
+                                "dimensions": {
+                                  "state": "active"
+                                }
+                              }
+                            ]
+                          }
+                        }
+                      ]
+                    }
+                    """;
+
+    final String cannedResponseForApplication3 =
+            """
+                    {
+                      "nodes": [
+                        {
+                          "hostname": "host-3.yahoo.com",
+                          "role": "role0",
+                          "node": {
+                            "timestamp": 1300,
+                            "metrics": [
+                              {
+                                "values": {
+                                  "cpu.util": 10,
+                                  "gpu.util": 13,
                                   "mem.util": 15,
+                                  "gpu.memory.used": 7.5,
+                                  "gpu.memory.total": 8,
                                   "disk.util": 20,
                                   "application_generation.last": 3,
                                   "in_service.last": 0