Add nvidia gpu topoloy score registry to node

docs/proposals/gpu-topo-policy.md

@@ -0,0 +1,197 @@
+# GPU topology scheduling strategy
+
+## Use Case
+
+### Node Select
+
+If multiple nodes meet the requirements, the minimum meeting node should be selected first::
+- `Node1`
+```json
+[
+  {
+    "uuid": "gpu0",
+    "score": {
+      "gpu1": "100",
+      "gpu2": "100",
+      "gpu3": "200"
+    }
+  },
+  {
+    "uuid": "gpu1",
+    "score": {
+      "gpu0": "100",
+      "gpu2": "200",
+      "gpu3": "100"
+    }
+  },
+  {
+    "uuid": "gpu2",
+    "score": {
+      "gpu0": "100",
+      "gpu1": "200",
+      "gpu3": "200"
+    }
+  },
+  {
+    "uuid": "gpu3",
+    "score": {
+      "gpu0": "200",
+      "gpu1": "100",
+      "gpu2": "200"
+    }
+  }
+]
+```
+
+
+- `Node2`
+```json
+[
+  {
+    "uuid": "gpu0",
+    "score": {
+      "gpu1": "100",
+      "gpu2": "100",
+      "gpu3": "200",
+      "gpu4": "200",
+      "gpu5": "200"
+    }
+  },
+  {
+    "uuid": "gpu1",
+    "score": {
+      "gpu0": "100",
+      "gpu2": "200",
+      "gpu3": "100",
+      "gpu4": "200",
+      "gpu5": "200"
+    }
+  },
+  {
+    "uuid": "gpu2",
+    "score": {
+      "gpu0": "100",
+      "gpu1": "200",
+      "gpu3": "200",
+      "gpu4": "200",
+      "gpu5": "200"
+    }
+  },
+  {
+    "uuid": "gpu3",
+    "score": {
+      "gpu0": "200",
+      "gpu1": "100",
+      "gpu2": "200",
+      "gpu4": "200",
+      "gpu5": "200"
+    }
+  },
+  {
+    "uuid": "gpu4",
+    "score": {
+      "gpu0": "200",
+      "gpu1": "100",
+      "gpu2": "200",
+      "gpu3": "200",
+      "gpu5": "200"
+    }
+  },
+  {
+    "uuid": "gpu5",
+    "score": {
+      "gpu0": "200",
+      "gpu1": "100",
+      "gpu2": "200",
+      "gpu3": "200",
+      "gpu4": "200"
+    }
+  }
+]
+```
+
+For example, the two nodes above should be given priority `Node1`.
+
+
+### Device Select, One Pod One Device
+
+1. When a Pod only needs one card, the card with the worst communication with other cards should be given priority if the video memory and computing power are sufficient, such as：
+- `Node1`
+```json
+[
+  {
+    "uuid": "gpu0",
+    "score": {
+      "gpu1": "100",
+      "gpu2": "100",
+      "gpu3": "200"
+    }
+  },
+  {
+    "uuid": "gpu1",
+    "score": {
+      "gpu0": "100",
+      "gpu2": "200",
+      "gpu3": "100"
+    }
+  },
+  {
+    "uuid": "gpu2",
+    "score": {
+      "gpu0": "100",
+      "gpu1": "200",
+      "gpu3": "200"
+    }
+  },
+  {
+    "uuid": "gpu3",
+    "score": {
+      "gpu0": "200",
+      "gpu1": "100",
+      "gpu2": "200"
+    }
+  }
+]
+```
+There are only four cards on this node, and `gpu0` and `gpu1` have the worst connectivity with other cards, so these two cards are preferred in single-card mode.
+
+### Device Select, One Pod More Than One Device
+1. When a Pod only needs multiple cards, if the video memory and computing power are sufficient, the card with the best communication with other cards should be given priority, such as：
+- `Node1`
+```json
+[
+  {
+    "uuid": "gpu0",
+    "score": {
+      "gpu1": "100",
+      "gpu2": "100",
+      "gpu3": "200"
+    }
+  },
+  {
+    "uuid": "gpu1",
+    "score": {
+      "gpu0": "100",
+      "gpu2": "200",
+      "gpu3": "100"
+    }
+  },
+  {
+    "uuid": "gpu2",
+    "score": {
+      "gpu0": "100",
+      "gpu1": "200",
+      "gpu3": "200"
+    }
+  },
+  {
+    "uuid": "gpu3",
+    "score": {
+      "gpu0": "200",
+      "gpu1": "100",
+      "gpu2": "200"
+    }
+  }
+]
+```
+There are only four cards on this node, and `gpu2` and `gpu3` have the best connectivity with other cards, so these two cards are given priority in multi-card mode.

pkg/device-plugin/nvidiadevice/nvinternal/plugin/register.go

@@ -33,7 +33,9 @@
 package plugin
 
 import (
+	"encoding/json"
 	"fmt"
+	"os"
 	"os/exec"
 	"strconv"
 	"strings"
@@ -185,8 +187,25 @@ func (plugin *NvidiaDevicePlugin) RegistrInAnnotation() error {
 		return err
 	}
 	encodeddevices := util.EncodeNodeDevices(*devices)
+	var data []byte
+	if os.Getenv("ENABLE_TOPOLOGY_SCORE") == "true" {
+		gpuScore, err := nvidia.CalculateGPUScore(util.GetDevicesUUIDList(*devices))
+		if err != nil {
+			klog.ErrorS(err, "calculate gpu topo score error")
+			return err
+		}
+		data, err = json.Marshal(gpuScore)
+		if err != nil {
+			klog.ErrorS(err, "marshal gpu score error.")
+			return err
+		}
+	}
+	klog.V(4).InfoS("patch nvidia  topo score to node", "hami.io/node-nvidia-score", string(data))
 	annos[nvidia.HandshakeAnnos] = "Reported " + time.Now().String()
 	annos[nvidia.RegisterAnnos] = encodeddevices
+	if len(data) > 0 {
+		annos[nvidia.RegisterGPUPairScore] = string(data)
+	}
 	klog.Infof("patch node with the following annos %v", fmt.Sprintf("%v", annos))
 	err = util.PatchNodeAnnotations(node, annos)